Transmission line corrosion state identification method based on deep learning hyperspectral analysis

Abstract

The invention discloses a transmission line corrosion state identification method based on deep learning hyperspectral analysis. The method comprises the following steps: S1, acquiring hyperspectral image data containing a transmission line and carrying out preprocessing; s2, carrying out corrosion area labeling on the hyperspectral preprocessing data set; s3, extracting multi-level edge features through multi-scale cavity convolution and a feature map pyramid unit; s4, constructing a local positive and negative sample pair by using the corrosion region and the adjacent non-corrosion region in the training data set, and calculating local comparison loss; s5, reasoning the hyperspectral preprocessing data set by adopting the optimized edge detection convolutional neural network; and S6, generating and outputting response disposal suggestions for different corrosion types and corrosion severity degrees.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of corrosion intelligent identification, and particularly relates to a transmission line corrosion state identification method based on deep learning hyperspectral analysis. BACKGROUND

[0002] In the fields of industry, transportation, ocean, aerospace, the corrosion problem of metal structure is always an important factor affecting the safety and service life of equipment. With the acceleration of industrialization process, the large use of various steel and alloy materials, the structure damage, equipment failure and safety hazards caused by corrosion are increasingly prominent. How to efficiently and accurately identify the corrosion area on the surface of metal structure and take timely maintenance and repair measures has become a key link to ensure the safe operation of industrial system. The traditional corrosion identification method mainly relies on manual inspection, and the inspection personnel observes, takes pictures, records and measures by hand to detect corrosion. This method not only consumes a lot of manpower and material resources, but also has low detection efficiency, and the result is highly dependent on the experience level of the detection personnel, so it is difficult to realize the standardization, intelligentization and automation of corrosion identification. In addition, manual detection is easily affected by environmental light, surface impurities, operation fatigue and other interference factors, and has problems such as missed detection and misjudgment, which is difficult to meet the corrosion detection needs of large area and complex structure.

[0003] In recent years, with the development of computer vision and image processing technology, automatic corrosion identification has gradually become a research hotspot. Related technologies mainly include traditional image processing algorithm-based methods and deep learning-based methods. Traditional image processing methods, such as edge detection, threshold segmentation and color analysis, can identify the corrosion area on the surface to a certain extent, but these methods have poor adaptability to different light conditions, complex backgrounds and various corrosion forms, are easily affected by noise, shadow and other interference, and have limited robustness and generalization ability. With the rise of deep learning, corrosion identification methods based on convolutional neural networks have attracted widespread attention. Convolutional neural networks have strong feature extraction capability and can automatically learn spatial and semantic features in images to improve the accuracy of corrosion identification. However, existing corrosion identification methods based on visible light images still have problems such as low recognition accuracy and limited generalization ability when dealing with complex scenes such as surface contamination, light changes and mixed corrosion types. In addition, single visible light information cannot fully reflect the changes of microstructure on the metal surface and corrosion products, which makes it difficult to accurately detect some early corrosion or hidden corrosion.

[0004] Hyperspectral imaging technology provides a new approach for corrosion identification due to its ability to acquire rich spectral information within a wide wavelength range. Hyperspectral images not only contain spatial information but also the spectral characteristics of each pixel, which can sensitively reflect subtle changes in material composition, corrosion products, and surface state. The corrosion detection method based on hyperspectral data can effectively distinguish different types of corrosion and surface contaminants, improving the sensitivity and accuracy of corrosion identification. Currently, some studies have attempted to apply hyperspectral imaging to metal corrosion detection by selecting wavebands, extracting features, and using classification algorithms to identify and segment corrosion areas. However, existing hyperspectral corrosion identification techniques still have many shortcomings. First, hyperspectral image data is large and has severe information redundancy. How to effectively preprocess and reduce dimensions while preserving key features related to corrosion is still a pressing problem. Second, existing feature extraction and classification methods cannot fully exploit the spatial-spectral joint features of hyperspectral data, resulting in suboptimal recognition results. Third, there is a lack of end-to-end intelligent identification and decision support systems tailored to actual application needs, making it difficult to achieve full-process automation from corrosion detection to maintenance recommendation output.

[0005] Therefore, how to provide a power transmission line corrosion state identification method based on deep learning hyperspectral analysis is a problem that those skilled in the art need to solve. SUMMARY

[0006] One object of the present application is to provide a power transmission line corrosion state identification method based on deep learning hyperspectral analysis. The present application combines local contrast learning algorithm, edge detection convolutional neural network, and hyperspectral imaging technology. Through deep analysis of multi-channel data of hyperspectral images, combined with multi-scale hollow convolution and feature pyramid network, high-precision pixel-level identification of power transmission line corrosion state is achieved. In the corrosion area boundary detection process, the system innovatively introduces a local contrast learning mechanism, dynamically constructs positive and negative sample pairs of corrosion and non-corrosion areas, and uses a joint loss optimization method to effectively improve the accuracy and robustness of corrosion boundary positioning. Combined with historical maintenance data and rule base, the system can intelligently generate response treatment recommendations for different types and severity of corrosion, with high intelligence and adaptability, significantly improving the automation level of power transmission line intelligent inspection and maintenance.

[0007] The power transmission line corrosion state identification method based on deep learning hyperspectral analysis according to the embodiment of the present application comprises the following steps:

[0008] S1, collect hyperspectral image data containing power transmission lines and perform preprocessing to obtain a hyperspectral preprocessing data set;

[0009] S2, label the corrosion area of the hyperspectral preprocessing data set to obtain a training data set with corrosion and non-corrosion area labels;

[0010] S3, adopting an edge detection convolutional neural network, taking a training data set as input, extracting multi-level edge features through a multi-scale hollow convolution and a feature map pyramid unit, and outputting a preliminary corrosion boundary feature map;

[0011] S4, based on a local contrast learning algorithm, constructing a local positive and negative sample pair by using a corrosion region and an adjacent non-corrosion region in the training data set in an edge detection convolutional neural network training process, calculating a local contrast loss, and optimizing the edge detection convolutional neural network together with a cross-entropy loss to output an optimized corrosion boundary feature map;

[0012] S5, using the optimized edge detection convolutional neural network to infer the hyperspectral pretreatment data set to obtain pixel-level corrosion recognition results;

[0013] S6, based on the pixel-level corrosion recognition results, combining historical maintenance data and a strategy rule library, generating and outputting response treatment suggestions for different corrosion types and corrosion severity.

[0014] Optionally, the hyperspectral image data specifically includes images of a 400-2500 nanometer waveband collected at 10 nanometer intervals, each pixel point corresponding to a reflectivity value of 212 spectral channels, and containing spatial coordinate information of the image and radiation of each pixel in the visible light, near-infrared and short-wave infrared waveband.

[0015] Optionally, the pixel-level corrosion recognition result specifically includes a corrosion site, a corrosion type, a corrosion area and a corrosion severity.

[0016] Optionally, the S2 specifically includes:

[0017] S21, screening out a region related to the power transmission line in the hyperspectral pretreatment data set to obtain a region to be labeled;

[0018] S22, referring to historical corrosion defect samples, performing detailed comparison on the region to be labeled, identifying a suspected corrosion area and a normal area, and generating a preliminary list of suspected corrosion areas;

[0019] S23, grouping and labeling the preliminary list of suspected corrosion areas by multiple labeling personnel, and independently giving corrosion and non-corrosion judgments;

[0020] S24, for the region where the labeling results differ, organizing all labeling personnel to centrally review and finally determine the labeling attention, and outputting the reviewed corrosion labeling result;

[0021] S25, one-to-one correspondence between the reviewed corrosion labeling result and the hyperspectral pretreatment data set, obtaining a training data set with corrosion and non-corrosion region labels.

[0022] Optionally, the S3 specifically includes:

[0023] S31. Construct an edge detection convolutional neural network with integrated residual connection and cross attention mechanism, normalize and suppress noise on the training dataset with eroded and non-eroded area labels, and input it into the edge detection convolutional neural network to obtain enhanced input data;

[0024] S32. Using the multi-branch structure of the edge detection convolutional neural network, the enhanced input data is input into the shallow branch for fine-grained texture edge capture and the deep branch for global structure analysis, and the cross-branch dynamic transmission of edge information is achieved through the feature interaction module between branches to obtain a fused edge feature map;

[0025] S33. A dilated convolution unit with a dynamically adjustable dilation rate is used for the fused edge feature map to achieve parallel extraction of multi-scale features. A spatially adaptive dilation rate selection mechanism is introduced to automatically match the optimal receptive field according to the changes in the corrosion area boundary. At the same time, parallel direction-sensitive convolution units are used to improve the feature response capability for long strip corrosion edges, outputting a multi-scale and direction-adaptive edge feature map.

[0026] S34, input the multi-scale and direction-adaptive edge feature map into the innovatively designed feature map pyramid unit, and adaptively fuse the edge responses of different scales and different semantic levels through the pyramid method to output a multi-level fused edge feature map;

[0027] S35, inputting the multi-level fused edge feature map into the contrast enhancement unit, further highlighting the significant features between the eroded boundary area and the background based on local context statistics and spatial gradient amplitude calculation, and obtaining a contrast-enhanced feature map;

[0028] S36. Perform pixel-level prediction on the contrast-enhanced feature map using the full convolution output unit to obtain a preliminary erosion boundary feature map.

[0029] Optionally, the S4 specifically includes:

[0030] S41. Based on the local contrast learning algorithm, the preliminary corrosion boundary feature map is extracted from the corrosion area and the adjacent non-corrosion area in the training data set at the pixel level. According to the local neighborhood partitioning method, positive sample pairs and negative sample pairs are constructed to generate local positive and negative sample pair label indexes.

[0031] S42. Using the local positive and negative sample pair label index, combined with the spatial information of the corrosion edge area and the local neighborhood characteristics, the positive sample pairs and the negative sample pairs are screened for high correlation, and a set of local positive and negative sample pairs with high similarity is generated;

[0032] S43, input the set of high-similarity local positive and negative sample pairs and the preliminary corrosion boundary feature map to a local contrast loss calculation unit, and calculate a local contrast loss value by using a local contrast learning algorithm;

[0033]

[0034] wherein L local is the local contrast loss value, N is the total number of local positive and negative sample pairs, f i is a feature vector of an i-th pixel in the preliminary corrosion boundary feature map, is a positive sample set corresponding to the i-th pixel, is a negative sample set corresponding to the i-th pixel, ‖f i -f j ‖2 is an Euclidean distance between the feature vector of the i-th pixel and a feature vector of a j-th pixel in the negative sample set, ‖f i -f k ‖2 is an Euclidean distance between the feature vector of the i-th pixel and a feature vector of a k-th pixel in the positive sample set, and a is a boundary interval hyperparameter used to adjust a minimum distance of separation between the positive and negative sample pairs, [·] + is a non-negative truncation operation, indicating taking a larger value between zero and the value in the parentheses, and ‖·‖2 is a two-norm;

[0035] S44, combine the local contrast loss value and a pixel-level cross-entropy loss value calculated based on the preliminary corrosion boundary feature map, perform normalized weighted fusion, and obtain a joint loss value;

[0036] S45, according to the joint loss value, perform end-to-end back propagation and network parameter optimization on the edge detection convolutional neural network, and obtain an optimized and trained edge detection convolutional neural network;

[0037] S46, input the training data set into the optimized and trained edge detection convolutional neural network, and output an optimized corrosion boundary feature map.

[0038] Optionally, the S5 specifically includes:

[0039] S51, perform normalization processing on the hyperspectral preprocessed data set, and obtain a normalized hyperspectral data set;

[0040] S52, input the normalized hyperspectral data set into the optimized and trained edge detection convolutional neural network, perform feature extraction, and obtain feature map data containing spatial and spectral information;

[0041] S53, perform multi-scale feature fusion on the feature map data, integrate feature information at different scales, and obtain a fused feature map;

[0042] S54, performing pixel-level classification prediction on the fused feature map to obtain a preliminary pixel classification result map, and determining each pixel as a corrosion region or a non-corrosion region;

[0043] S55, post-processing the preliminary pixel classification result map, and optimizing by using a morphological processing method to obtain a pixel segmentation map with noise removed;

[0044] S56, outputting the pixel segmentation map with noise removed as a pixel-level corrosion identification result.

[0045] Optionally, the S6 specifically comprises:

[0046] S61, performing feature statistics on the pixel-level corrosion identification result, generating corrosion feature statistical data based on shape, area and distribution information of the corrosion region;

[0047] S62, correlating the corrosion feature statistical data with historical maintenance data, matching historical cases and supplementing relevant parameters to obtain correlated corrosion case data;

[0048] S63, combining the correlated corrosion case data, performing rule retrieval and condition judgment according to a strategy rule library to obtain preliminary response treatment suggestion data;

[0049] S64, optimizing and adjusting the preliminary response treatment suggestion data in combination with a current corrosion severity determination result to generate response treatment suggestions for different corrosion types and corrosion severity.

[0050] The present application has the following advantages:

[0051] The present application establishes a complete corrosion intelligent identification and response treatment suggestion system by fusing hyperspectral imaging and deep learning technology. Compared with the prior art, the present application can automatically detect the corrosion region on the surface of the metal structure with high precision and multiple types, which not only significantly improves the accuracy and robustness of corrosion identification, but also greatly reduces the workload and subjective error of manual detection. Through multi-dimensional data preprocessing of hyperspectral images and optimization of feature extraction by deep convolutional neural network, the system can fully exploit the spatial and spectral features of the corrosion region, and effectively identify complex environments, different corrosion types and early corrosion phenomena. Further, the present application combines pixel-level corrosion identification results with historical maintenance data and a strategy rule library, which not only automatically analyzes corrosion types and corrosion severity, but also intelligently outputs targeted maintenance and treatment suggestions according to actual working conditions and historical experience, greatly improving the scientificity and automation level of corrosion management and decision-making. Overall, the present application realizes an integrated automatic process of corrosion detection, analysis and maintenance suggestion, significantly improves the efficiency and intelligent level of industrial equipment safety management, and has a wide application prospect and significant engineering value. BRIEF DESCRIPTION OF DRAWINGS

[0052] The accompanying drawings are included to provide a further understanding of the application and are incorporated in and constitute a part of this specification, illustrate embodiments of the application and are meant to explain the present application and are not intended to limit the application. In the drawings:

[0053] Fig. 1 A flow chart of the power transmission line corrosion state recognition method based on deep learning hyperspectral analysis proposed by the present application;

[0054] Fig. 2 A schematic diagram of the power transmission line corrosion state recognition method based on deep learning hyperspectral analysis proposed by the present application;

[0055] Fig. 3 A data flow diagram of the power transmission line corrosion state recognition method based on deep learning hyperspectral analysis proposed by the present application. DETAILED DESCRIPTION

[0056] The present application will now be described in further detail with reference to the drawings. These drawings are simplified schematic diagrams which show the basic structure of the present application in a schematic manner only, and thus only show the components relevant to the present application.

[0057] REFERENCE Figs. 1-3 The power transmission line corrosion state recognition method based on deep learning hyperspectral analysis comprises the following steps:

[0058] S1, collecting hyperspectral image data containing power transmission lines and performing preprocessing to obtain a hyperspectral preprocessing data set;

[0059] S2, labeling the corrosion area of the hyperspectral preprocessing data set to obtain a training data set with corrosion and non-corrosion area labels;

[0060] S3, using an edge detection convolutional neural network, taking the training data set as input, extracting multi-level edge features through multi-scale hollow convolution and feature pyramid unit, and outputting a preliminary corrosion boundary feature map;

[0061] S4, based on a local contrast learning algorithm, in the edge detection convolutional neural network training process, using the corrosion area and adjacent non-corrosion area in the training data set to construct a local positive and negative sample pair, calculating the local contrast loss, and optimizing the edge detection convolutional neural network together with the cross-entropy loss, and outputting the optimized corrosion boundary feature map;

[0062] S5, using the optimized edge detection convolutional neural network to infer the hyperspectral preprocessing data set to obtain pixel-level corrosion recognition results;

[0063] S6, based on the pixel-level corrosion recognition result, combined with historical maintenance data and policy rule library, response treatment suggestions for different corrosion types and corrosion severity are generated and output.

[0064] The application realizes pixel-level accurate recognition of corrosion areas by fusing hyperspectral imaging and edge detection convolutional neural network combined with local contrast learning algorithm, and intelligently generates response treatment suggestions through historical maintenance data and policy rule library, which significantly improves the automation and decision-making scientificity of transmission line corrosion detection.

[0065] In the embodiment, the hyperspectral image data specifically includes images of 400-2500 nanometer waveband collected at 10 nanometer intervals, each pixel point corresponds to reflectivity values of 212 spectral channels, and includes spatial coordinate information of the image and radiation of each pixel in the visible light, near infrared and short wave infrared waveband.

[0066] The application adopts a hyperspectral image collection method of 400-2500 nanometer, 10 nanometer interval, obtains 212 spectral channel reflectivity and spatial coordinates and multi-band radiation information per pixel, realizes multi-dimensional accurate characterization of physical and chemical characteristics of corrosion areas, and greatly improves the corrosion type distinguishing and early identification capability.

[0067] In the embodiment, the pixel-level corrosion recognition result specifically includes corrosion site, corrosion type, corrosion area and corrosion severity.

[0068] The application realizes comprehensive and detailed quantitative analysis of corrosion conditions by outputting pixel-level corrosion recognition results including corrosion site, corrosion type, corrosion area and corrosion severity, which is convenient for accurate positioning of risk areas and provides scientific basis for subsequent operation and maintenance decision and targeted treatment measures.

[0069] In the embodiment, S2 specifically includes:

[0070] S21, the region related to the transmission line in the hyperspectral preprocessing data set is screened out, and a region to be labeled is obtained;

[0071] S22, referring to historical corrosion defect samples, the region to be labeled is compared in detail, suspected corrosion areas and normal areas are identified, and a preliminary list of suspected corrosion areas is generated;

[0072] S23, the preliminary list of suspected corrosion areas is grouped and labeled by multiple labeling personnel, and each independently gives corrosion and non-corrosion judgment;

[0073] S24, for the region with different labeling results, all labeling personnel are organized to concentrate on reviewing and finally determine the labeling result, and output the reviewed corrosion labeling result;

[0074] S25, correspond the reviewed corrosion annotation result to the hyperspectral pretreatment data set one by one, obtain the training data set with corrosion and non-corrosion area label.

[0075] The application ensures high accuracy and consistency of corrosion and non-corrosion area labels through historical defect sample comparison, multi-person grouping independent annotation and centralized review mechanism, significantly improves the annotation quality of the training data set, provides stable and reliable data foundation for subsequent model training, and improves the generalization ability and practical value of the corrosion detection model.

[0076] In the embodiment, the S3 specifically comprises:

[0077] S31, an edge detection convolutional neural network integrating residual connection and cross attention mechanism is constructed, the training data set with corrosion and non-corrosion area label is normalized and noise suppressed, and then input into the edge detection convolutional neural network to obtain enhanced input data;

[0078] S32, the multi-branch structure of the edge detection convolutional neural network is used to input the enhanced input data into a shallow branch for capturing fine-grained texture edges and a deep branch for global structure analysis, and the cross-branch dynamic transmission of edge information is realized through a feature interaction module between the branches to obtain a fused edge feature map;

[0079] S33, a dilated convolution unit with a dynamically adjustable dilated rate is used for the fused edge feature map to realize parallel extraction of multi-scale features, a spatial adaptive dilated rate selection mechanism is introduced to automatically match the optimal receptive field according to the corrosion area boundary change, and at the same time, a direction-sensitive convolution unit in parallel is used to improve the feature response ability to long strip-shaped corrosion edges, and a multi-scale and direction-adaptive edge feature map is output;

[0080] S34, the multi-scale and direction-adaptive edge feature map is input into an innovatively designed feature pyramid unit to pyramid adaptively fuse edge responses of different scales and different semantic levels, and a multi-level fused edge feature map is output;

[0081] S35, the multi-level fused edge feature map is input into a contrast enhancement unit based on local context statistics and spatial gradient amplitude calculation to further highlight the salient features between the corrosion boundary area and the background, and a contrast-enhanced feature map is obtained;

[0082] S36, a full convolution output unit is used for pixel-level prediction of the contrast-enhanced feature map to obtain a preliminary corrosion boundary feature map.

[0083] The application realizes multi-scale, direction adaptive and efficient extraction of fine-grained corrosion edges by integrating residual connection, cross attention, multi-branch feature interaction and dynamic hollow convolution structures, and effectively improves the accuracy of preliminary corrosion detection and the robustness in complex scenes by means of feature map pyramid and contrast enhancement to highlight the corrosion boundary.

[0084] In the embodiment, the S4 specifically comprises:

[0085] S41, based on a local contrast learning algorithm, pixel-level extraction is performed on the preliminary corrosion boundary feature map and the corrosion region and adjacent non-corrosion region in the training data set, a positive sample pair and a negative sample pair are constructed according to a local neighborhood division method, and a local positive and negative sample pair label index is generated;

[0086] S42, using the local positive and negative sample pair label index, combining the spatial information of the corrosion edge region and the local neighborhood features, the positive sample pair and the negative sample pair are high-correlation filtered to generate a high-similarity local positive and negative sample pair set;

[0087] S43, inputting the high-similarity local positive and negative sample pair set and the preliminary corrosion boundary feature map into a local contrast loss calculation unit, and calculating a local contrast loss value by using a local contrast learning algorithm:

[0088]

[0089] Wherein, L local is the local contrast loss value, N is the total number of local positive and negative sample pairs, f i is the feature vector of the i-th pixel in the preliminary corrosion boundary feature map, is the positive sample set corresponding to the i-th pixel, is the negative sample set corresponding to the i-th pixel, ‖f i -f j ‖2 is the Euclidean distance between the feature vector of the i-th pixel and the j-th pixel in the negative sample set, ‖f i -f k ‖2 is the Euclidean distance between the feature vector of the i-th pixel and the k-th pixel in the positive sample set, and alpha is a boundary interval hyperparameter, used to adjust the minimum distance between the positive and negative sample pairs, [·] + is a non-negative truncation operation, indicating taking the larger value of zero and the value in the parentheses, and ‖·‖2 is the two-norm;

[0090] S44, combining the local contrast loss value and the pixel-level cross-entropy loss value calculated based on the preliminary corrosion boundary feature map, performing normalized weighted fusion to obtain a joint loss value;

[0091] S45, according to the joint loss value, end-to-end back propagation and network parameter optimization are carried out on the edge detection convolutional neural network, and an optimized and trained edge detection convolutional neural network is obtained.

[0092] S46, the training data set is input into the optimized and trained edge detection convolutional neural network, and an optimized corrosion boundary feature map is output.

[0093] The application introduces a local contrast learning algorithm, combines high-similarity positive and negative sample pairs and spatial neighborhood features, jointly optimizes the network by local contrast loss and pixel-level cross-entropy loss, significantly enhances the discrimination ability of corrosion boundaries, and effectively improves the precision of corrosion detection under complex background and the sensitivity of the model to small corrosion areas.

[0094] In the embodiment, the S5 specifically includes:

[0095] S51, the hyperspectral preprocessed data set is normalized to obtain a normalized hyperspectral data set;

[0096] S52, the normalized hyperspectral data set is input into the optimized and trained edge detection convolutional neural network for feature extraction, and a feature map data containing spatial and spectral information is obtained;

[0097] S53, multi-scale feature fusion is performed on the feature map data, and feature information at different scales is integrated to obtain a fused feature map;

[0098] S54, pixel-level classification prediction is performed on the fused feature map to obtain a preliminary pixel classification result map, and each pixel is determined as a corrosion area or a non-corrosion area;

[0099] S55, the preliminary pixel classification result map is post-processed, and a morphological processing method is used for optimization to obtain a pixel segmentation map with noise removed;

[0100] S56, the pixel segmentation map with noise removed is output as a pixel-level corrosion recognition result.

[0101] The application realizes accurate segmentation of hyperspectral pixel-level corrosion areas by using the multi-scale feature fusion and morphological post-processing method, effectively removes noise and misjudgment, improves the accuracy and robustness of corrosion recognition, and provides reliable data basis and technical support for subsequent corrosion state evaluation and fine operation and maintenance.

[0102] In the embodiment, the S6 specifically includes:

[0103] S61, feature statistics are performed on the pixel-level corrosion recognition result, corrosion feature statistical data are generated based on shape, area and distribution information of the corrosion area;

[0104] S62, correlate the corrosion feature statistical data with historical maintenance data, match historical cases and supplement relevant parameters, and obtain associated corrosion case data;

[0105] S63, in combination with the associated corrosion case data, rule retrieval and condition judgment are performed according to the strategy rule base, and preliminary response treatment suggestion data are obtained;

[0106] S64, the preliminary response treatment suggestion data are optimized and adjusted in combination with the current corrosion severity determination result, and response treatment suggestions for different corrosion types and corrosion severities are generated.

[0107] The present application realizes the individualization optimization of response treatment suggestions by intelligent comparison of corrosion area feature statistical data and historical maintenance cases, in combination with the strategy rule base and the corrosion severity, improves the pertinence and scientificity of treatment suggestions, and effectively supports the accurate decision and efficient operation and maintenance management of corrosion defects in complex scenarios.

[0108] Embodiment 1:

[0109] In order to verify the feasibility of the present application in implementation, the present application is applied to the "220kV high-voltage transmission line corrosion detection and intelligent operation and maintenance" project carried out by A City Power Transmission Company. The line is 56 kilometers long, passes through hills, plains and part of urban residential areas, and is an important part of the main network of A City. For a long time, due to humid climate, high salt fog and industrial emissions, corrosion frequently occurs in line equipment, especially in iron towers, cross arms, connecting parts and other parts. The traditional inspection method relies on manual visual inspection, has low efficiency, strong subjectivity, and is difficult to accurately identify early corrosion and hidden parts, resulting in that some corrosion is not discovered in time, affecting the safe operation of the line.

[0110] In this project, the operation and maintenance team adopts the hyperspectral remote sensing and intelligent edge detection integrated corrosion identification and response method proposed by the present application. First, a high-resolution unmanned aerial vehicle is used to carry a hyperspectral imaging device to take pictures of the whole line every month, and about 320GB of original hyperspectral image data is obtained in a single task. Through the data preprocessing process of the present application, irrelevant areas are excluded, only areas related to the transmission line are retained, and suspicious areas to be labeled are automatically selected, greatly reducing the burden of manual screening.

[0111] In the corrosion area manual labeling stage, the system automatically compares the historical corrosion defect sample database and preliminarily classifies the suspicious areas in each collection data. The labeling work is independently completed by 5 experienced technicians, and the consistency of the labeling results is ensured through grouping and centralized review. Taking one data collection as an example, the system preliminarily selects 89 suspected corrosion areas, and finally determines 71 corrosion areas and 18 normal areas after manual confirmation. Compared with the traditional manual point-by-point labeling of the whole image, the labeling efficiency is improved by about 67%, and the labeling difference rate is reduced to 1.4%.

[0112] The data set with accurate labels is input into the edge detection convolutional neural network with designed integrated residual connection and cross attention mechanism. After normalization, noise suppression and multi-branch feature extraction, the system can automatically focus on the tower cross arm weld, bolt connection and other easy-to-corrode details. The introduction of multi-scale hollow convolution and direction-sensitive convolution enables long strip, point and sheet corrosion boundaries to be accurately captured. After feature pyramid and contrast enhancement processing, the pixel-level corrosion boundary recognition rate is increased from 78.1% of the traditional method to 91.5%, and the recall rate of small area corrosion area reaches 86.7%, which is better than the previous single CNN-based scheme.

[0113] After pixel-level corrosion recognition, the local contrast learning algorithm of the present application is used to compare the pixels of the corrosion area and the surrounding non-corrosion area, which significantly improves the clarity and accuracy of boundary recognition. The corrosion boundary error using the method of the present application is reduced to an average of 1.3 pixels, which is significantly better than the 3.6 pixels of the traditional method, and the F1 score in the complex background is stable at 0.91 or above.

[0114] Table 1 Comparison of optimization effect of power transmission line corrosion state recognition based on deep learning hyperspectral analysis

[0115]

[0116] Table 1 shows that the method of the present application significantly improves the accuracy, processing efficiency and intelligent level of response of corrosion recognition during the implementation of the present project. The average time for a single full-line corrosion detection task is reduced from 12 days to 3 days, the manual participation is reduced by 70%, the corrosion false positive rate is reduced from 15.6% to 4.8%, and the undetected rate is reduced from 9.3% to 1.2%. Compared with the same period in the past three years, the failure rate of the line caused by corrosion is reduced by 68%, the operation and maintenance cost is saved by about 184,000 yuan, and the average service life of the equipment is estimated to be increased by 1.7 years. The operation and maintenance personnel feedback that the disposal suggestions output by the system have strong pertinence and operability, which greatly reduces the work pressure of the front-line personnel.

[0117] The above describes only the preferred specific embodiments of the present application, but the protection scope of the present application is not limited thereto, and any skilled person in the art can make equivalent substitutions or changes within the technical scope disclosed by the present application according to the technical solutions and inventive concepts of the present application, which should be covered within the protection scope of the present application.

Claims

1. A method for identifying the corrosion status of transmission lines based on deep learning hyperspectral analysis, characterized in that: The steps include: S1, collecting and preprocessing hyperspectral image data containing transmission lines to obtain a hyperspectral preprocessed dataset; S2. Label the corrosion regions of the hyperspectral preprocessed dataset to obtain a training dataset with labels of corrosion and non-corrosion regions; S3, using edge detection convolutional neural network, taking the training data set as input, extracting multi-level edge features through multi-scale hole convolution and feature map pyramid unit, and outputting preliminary erosion boundary feature map; S4. Based on the local contrast learning algorithm, during the training process of the edge detection convolutional neural network, the corroded area and the adjacent non-corroded area in the training dataset are used to construct local positive and negative sample pairs, and the local contrast loss is calculated. The local contrast loss is then used together with the cross entropy loss to optimize the edge detection convolutional neural network and output the optimized corrosion boundary feature map. S5. Use the optimized edge detection convolutional neural network to infer the hyperspectral preprocessed dataset and obtain pixel-level corrosion recognition results; S6. Based on the pixel-level corrosion identification results, combined with historical maintenance data and policy rule base, generate and output response and disposal recommendations for different corrosion types and corrosion severity.

2. The method for identifying the corrosion status of a transmission line based on deep learning hyperspectral analysis according to claim 1 is characterized in that: The hyperspectral image data specifically includes images in the 400 nm to 2500 nm band collected at 10 nm intervals. Each pixel corresponds to the reflectance value of 212 spectral channels, and contains the spatial coordinate information of the image, as well as the radiation of each pixel in the visible light, near-infrared and short-wave infrared bands.

3. The method for identifying the corrosion status of a transmission line based on deep learning hyperspectral analysis according to claim 1 is characterized in that: The pixel-level corrosion identification results specifically include corrosion location, corrosion type, corrosion area and corrosion severity.

4. The method for identifying the corrosion status of a transmission line based on deep learning hyperspectral analysis according to claim 1 is characterized in that: The S2 specifically includes: S21, filtering out the areas related to the transmission lines in the hyperspectral preprocessing data set to obtain areas to be labeled; S22. Refer to historical corrosion defect samples to conduct detailed comparison of the marked areas, identify suspected corrosion areas and normal areas, and generate a preliminary list of suspected corrosion areas; S23. Organize multiple annotators to group and annotate the preliminary list of suspected corrosion areas, and each person independently gives a judgment of corrosion or non-corrosion; S24. For areas where there are discrepancies in the annotation results, organize all annotation personnel to conduct a centralized review and finalize the annotation opinions, and output the reviewed corrosion annotation results; S25. The verified corrosion annotation results are matched one-to-one with the hyperspectral preprocessing dataset to obtain a training dataset with labels of corrosion and non-corrosion areas.

5. The method for identifying the corrosion status of a transmission line based on deep learning hyperspectral analysis according to claim 1 is characterized in that: The S3 specifically includes: S31. Construct an edge detection convolutional neural network with integrated residual connection and cross attention mechanism, normalize and suppress noise on the training dataset with eroded and non-eroded area labels, and input it into the edge detection convolutional neural network to obtain enhanced input data; S32. Using the multi-branch structure of the edge detection convolutional neural network, the enhanced input data is input into the shallow branch for fine-grained texture edge capture and the deep branch for global structure analysis, and the cross-branch dynamic transmission of edge information is achieved through the feature interaction module between branches to obtain a fused edge feature map; S33. A dilated convolution unit with a dynamically adjustable dilation rate is used for the fused edge feature map to achieve parallel extraction of multi-scale features. A spatially adaptive dilation rate selection mechanism is introduced to automatically match the optimal receptive field according to the changes in the corrosion area boundary. At the same time, parallel direction-sensitive convolution units are used to improve the feature response capability for long strip corrosion edges, outputting a multi-scale and direction-adaptive edge feature map. S34, input the multi-scale and direction-adaptive edge feature map into the innovatively designed feature map pyramid unit, and adaptively fuse the edge responses of different scales and different semantic levels through the pyramid method to output a multi-level fused edge feature map; S35, inputting the multi-level fused edge feature map into the contrast enhancement unit, further highlighting the significant features between the eroded boundary area and the background based on local context statistics and spatial gradient amplitude calculation, and obtaining a contrast-enhanced feature map; S36. Perform pixel-level prediction on the contrast-enhanced feature map using the full convolution output unit to obtain a preliminary erosion boundary feature map.

6. The method for identifying the corrosion status of a transmission line based on deep learning hyperspectral analysis according to claim 1 is characterized in that: The S4 specifically includes: S41. Based on the local contrast learning algorithm, the preliminary corrosion boundary feature map is extracted from the corrosion area and the adjacent non-corrosion area in the training data set at the pixel level. According to the local neighborhood partitioning method, positive sample pairs and negative sample pairs are constructed to generate local positive and negative sample pair label indexes. S42. Using the local positive and negative sample pair label index, combined with the spatial information of the corrosion edge area and the local neighborhood characteristics, the positive sample pairs and the negative sample pairs are screened for high correlation, and a set of local positive and negative sample pairs with high similarity is generated; S43. Input the high-similarity local positive and negative sample pair set and the preliminary erosion boundary feature map into the local contrast loss calculation unit, and use the local contrast learning algorithm to calculate the local contrast loss value: Among them, L local is the local contrast loss value, N is the total number of local positive and negative sample pairs, f i is the feature vector of the i-th pixel in the preliminary erosion boundary feature map, is the set of positive samples corresponding to the i-th pixel, is the set of negative samples corresponding to the i-th pixel, ‖f i -f j ‖2 is the Euclidean distance between the feature vector of the i-th pixel and the j-th pixel in the negative sample set, ‖f i -f k ‖2 is the Euclidean distance between the feature vector of the i-th pixel and the k-th pixel in the positive sample set, α is the boundary interval hyperparameter used to adjust the minimum distinguishing distance between positive and negative sample pairs, [·] + is a non-negative truncation operation, which means taking the larger value of zero and the value in the brackets, and ‖·‖2 is the two-norm; S44, combining the local contrast loss value and the pixel-level cross entropy loss value calculated based on the preliminary erosion boundary feature map, performing normalized weighted fusion to obtain a joint loss value; S45. Perform end-to-end backpropagation and network parameter optimization on the edge detection convolutional neural network based on the joint loss value to obtain an optimized trained edge detection convolutional neural network; S46. Input the training data set into the optimized trained edge detection convolutional neural network, and output the optimized corrosion boundary feature map.

7. The method for identifying the corrosion status of a transmission line based on deep learning hyperspectral analysis according to claim 1 is characterized in that: The S5 specifically includes: S51, performing normalization processing on the hyperspectral preprocessing data set to obtain a normalized hyperspectral data set; S52, inputting the normalized hyperspectral data set into an optimized and trained edge detection convolutional neural network to perform feature extraction to obtain feature map data containing spatial and spectral information; S53, performing multi-scale feature fusion on the feature map data, integrating feature information at different scales to obtain a fused feature map; S54, performing pixel-level classification prediction on the fused feature map to obtain a preliminary pixel classification result map, and determining each pixel as a corrosion area or a non-corrosion area; S55, post-processing the preliminary pixel classification result map, optimizing it using a morphological processing method, and obtaining a pixel segmentation map with noise removed; S56: Output the pixel segmentation map after removing the noise as a pixel-level corrosion recognition result.

8. The method for identifying the corrosion status of a transmission line based on deep learning hyperspectral analysis according to claim 1 is characterized in that: The S6 specifically includes: S61, performing feature statistics on the pixel-level corrosion identification results, and generating corrosion feature statistics based on the shape, area, and distribution information of the corrosion area; S62, correlating the corrosion characteristic statistical data with the historical maintenance data, matching the historical cases and supplementing the relevant parameters to obtain the correlated corrosion case data; S63. Combine the associated corrosion case data, perform rule retrieval and condition judgment according to the strategy rule library, and obtain preliminary response and disposal suggestion data; S64. Optimize and adjust the preliminary response and disposal recommendation data in combination with the current corrosion severity determination results to generate response and disposal recommendations for different corrosion types and corrosion severities.

Images