Transformer substation intelligent detection method and satellite image processing system

Abstract

The invention relates to the technical field of intelligent operation inspection of power transmission lines, in particular to an intelligent substation detection method and a satellite image processing system.The method comprises the steps that target data information is collected, and a power station detection data set is constructed by obtaining a historical multi-dimensional feature data set; a multi-level feature extraction network is established to perform multi-target feature extraction on target data information, a cavity convolution module is set to perform feature fusion on multi-target feature mapping, and a power station comprehensive detection model is constructed based on the multi-level feature extraction network and the cavity convolution module. According to the power station comprehensive detection method and device, the problems that a traditional detection method is prone to being interfered by factors such as weather and terrains in a complex environment, the recognition effect is unstable, and the precision is low are effectively solved; and the fusion capability of the multi-level features is effectively improved, so that high-precision identification of the transformer substation and equipment thereof is ensured.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent operation and maintenance of transmission lines, and particularly to an intelligent detection method for substations and a satellite image processing system. Background Art

[0002] With the rapid development of technology, electricity has gradually become an indispensable part of modern life and industrial production. The existence of substations has greatly improved the flexibility and reliability of the power system. By adjusting the voltage levels in different regions, substations can flexibly allocate power resources according to actual needs, effectively avoiding large-scale power outages caused by supply-demand mismatches. This function is crucial for ensuring the basic electricity consumption of residents in urban and remote areas. Therefore, ensuring the safe and stable operation of substations is particularly important.

[0003] However, currently, substations are usually located in outdoor environments, where there may be various obstacles such as trees and buildings around them. At the same time, they are also affected by factors such as light changes and weather conditions. Existing detection methods often have difficulty accurately distinguishing substations from other similar objects, resulting in inaccurate data information and posing great safety hazards to the operation and maintenance of transmission lines.

[0004] The information disclosed in this background art section is only intended to deepen the understanding of the overall background art of the present disclosure, and should not be regarded as an admission or any form of implication that this information constitutes the prior art known to those skilled in the art. Summary of the Invention

[0005] The present invention provides an intelligent detection method for substations and a satellite image processing system, which can effectively solve the problems in the background art.

[0006] In order to achieve the above object, the technical solution adopted by the present invention is: An intelligent detection method for substations, the method comprising: Collecting target data information, obtaining a historical multi-dimensional feature dataset, and constructing a power station detection dataset according to the historical multi-dimensional feature dataset; Establishing a multi-level feature extraction network, performing multi-target feature extraction on the target data information according to the multi-level feature extraction network to obtain a multi-target feature map, and setting a dilated convolution module to perform feature fusion on the multi-target feature map to obtain a comprehensive power station feature; Constructing a comprehensive power station detection model based on the multi-level feature extraction network and the dilated convolution module, and training the comprehensive power station detection model according to the power station detection dataset; Performing power station detection on the target data information according to the trained comprehensive power station detection model to obtain a power station detection result.

[0007] Further, perform multi-object feature extraction on the target data information according to the multi-level feature extraction network to obtain a multi-object feature map, including: Construct a power station feature database according to the historical multi-dimensional feature dataset, perform feature extraction on the target data information according to the power station feature database to obtain a number of preliminary feature maps, and the preliminary feature maps include a number of low-level feature maps and a number of high-level feature maps; Based on a number of the preliminary feature maps, establish residual skip connection paths respectively, and the residual skip connection paths are used to directly link the low-level feature maps and the high-level feature maps; Set cross-layer feature transfer weights according to a number of the preliminary feature maps, and set skip connection parameters according to the residual skip connection paths. Dynamically adjust the cross-layer feature transfer weights based on the skip connection parameters to obtain a multi-object feature map.

[0008] Further, dynamically adjusting the cross-layer feature transfer weights based on the skip connection parameters includes: Calculate the feature similarity between the low-level feature map and the high-level feature map according to the power station feature database, and set the initial cross-layer feature transfer weight according to the feature similarity; Calculate a cross-layer influence factor according to the initial cross-layer feature transfer weight, and correct the skip connection parameters according to the cross-layer influence factor; Re-distribute the cross-layer feature transfer weights based on the corrected skip connection parameters.

[0009] Further, perform power station detection on the target data information according to the trained power station comprehensive detection model, including: Extract multi-level feature maps of the target data information, where the low-level feature maps contain texture details of power equipment, and the high-level feature maps contain geometric contours of the power equipment; Perform feature weighted integration on the low-level feature maps and the high-level feature maps to obtain an enhanced feature map; Perform local area scanning on the enhanced feature map to determine the equipment area, and extract the physical features and spatial coordinates of the power equipment in the equipment area; Obtain the historical equipment relationship of the power station, perform consistency verification on the power equipment in the equipment area according to the historical equipment relationship of the power station, and output the power station facility positioning information.

[0010] Further, constructing a power station comprehensive detection model includes: Match and compare the comprehensive power station features according to the power station facility positioning information to obtain the spatial correspondence between the power station facility positioning information and the comprehensive power station features; If the power station facility location information does not match the comprehensive power station characteristics, calculate the feature contribution weight based on the power station facility location information, and correct the category distribution of the comprehensive power station characteristics according to the feature contribution weight to match the power station facility location information. The feature contribution weight is allocated according to the number of pixels in the regional category; If the power station facility location information matches the comprehensive power station characteristics, fuse the power station facility location information and the comprehensive power station characteristics to construct a power station comprehensive detection model.

[0011] Further, obtain the power station detection result and perform image optimization on the power station detection result, including: Obtain the misdetection image and the standard image according to the power station detection result, perform feature reshaping on the misdetection image, and check the standard image; Perform multiple image detail optimizations on the misdetection image to obtain several optimized comparison images with different degrees; Compare several of the optimized comparison images with the misdetection image respectively, and mark the feature difference regions of the misdetection image; Merge the feature difference regions with the power station detection data set to update the power station detection data set.

[0012] Further, the checking of the standard image includes: Collect the historical ledger coordinates, extract the centroid coordinates of the power station area according to the standard image, compare the historical ledger coordinates with the centroid coordinates of the power station area, and obtain the plane distance difference; Set an offset threshold according to the historical equipment relationship of the power station. If the plane distance difference exceeds the offset threshold, perform secondary verification on the standard image to obtain the corrected offset; Correct the standard image according to the corrected offset and check it again until the plane distance difference meets the offset threshold.

[0013] Further, train the power station comprehensive detection model according to the power station detection data set, including: Clean the historical multi-dimensional feature data set to obtain the original training set; Perform data information processing on several historical multi-dimensional feature data sets respectively according to the original training set to construct a power station detection data set. The data information processing includes target recognition and manual correction; Divide the power station detection data set into a power station training set and a power station test set, train the power station comprehensive detection model according to the power station training set, and verify the power station comprehensive detection model according to the power station test set.

[0014] A satellite image processing system, the system comprising: A data information construction module, which uses a remote sensing satellite to collect target data information, obtains a historical multi-dimensional feature dataset, and constructs a power station detection dataset according to the historical multi-dimensional feature dataset; A feature extraction and fusion module, which establishes a multi-level feature extraction network, performs multi-target feature extraction on the target data information according to the multi-level feature extraction network, obtains a multi-target feature map, and sets a dilated convolution module to perform feature fusion on the multi-target feature map to obtain a comprehensive power station feature; An identification model construction and training module, which constructs a comprehensive power station detection model based on the multi-level feature extraction network and the dilated convolution module, and trains the comprehensive power station detection model according to the power station detection dataset; A power station identification and output module, which performs power station detection on the target data information according to the trained comprehensive power station detection model to obtain a power station detection result.

[0015] Furthermore, the feature extraction and fusion module includes: A power station feature extraction and analysis unit, which constructs a power station feature database according to the historical multi-dimensional feature dataset, performs feature extraction on the target data information according to the power station feature database, obtains a number of preliminary feature maps, and the preliminary feature maps include a number of low-level feature maps and a number of high-level feature maps; A residual connection feature transfer unit, which respectively establishes residual skip connection paths based on a number of the preliminary feature maps, and the residual skip connection paths are used to directly link low-level feature maps and high-level feature maps; A cross-layer weight dynamic adjustment unit, which sets cross-layer feature transfer weights according to a number of the preliminary feature maps, sets skip connection parameters according to the residual skip connection paths, and dynamically adjusts the cross-layer feature transfer weights based on the skip connection parameters to obtain a multi-target feature map.

[0016] Through the technical solution of the present invention, the following technical effects can be achieved: Effectively solves the problem that traditional detection methods are easily interfered by factors such as weather and terrain in complex environments, resulting in unstable recognition effects and low accuracy. By combining a multi-level feature extraction network and a dilated convolution module, the fusion ability of multi-level features is effectively improved, thus ensuring high-precision recognition of substations and their equipment. At the same time, the training and optimization of the deep learning model enable the present invention to operate stably in complex environments, significantly improving the accurate positioning of power station facilities and meeting the high-precision recognition requirements.

[0017] The above description is only an overview of the technical solution of the present application. In order to better understand the technical means of the present application, it can be implemented according to the content of the description. In order to make the above and other objects, features and advantages of the present application more obvious and understandable, the following specific embodiments of the present application are given. Brief Description of the Drawings

[0018] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are only some embodiments recorded in the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0019] Figure 1 It is a schematic flow chart of a substation intelligent detection method; Figure 2 It is a schematic flow chart of multi-target feature mapping extraction; Figure 3 It is a schematic flow chart of dynamic adjustment of cross-layer feature transfer weights; Figure 4 It is a schematic diagram of optimizing the substation detection results; Figure 5 It is a schematic diagram of a satellite image processing system. Detailed Description of the Embodiments

[0020] The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments.

[0021] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the technical field to which the present invention belongs. The terms used in the description of the present invention are only for the purpose of describing specific embodiments, and are not intended to limit the present invention. The term "and / or" used herein includes any and all combinations of one or more of the related listed items.

[0022] Embodiment 1; As Figure 1 shown, the present application provides a substation intelligent detection method, and the method includes: S100: Use a remote sensing satellite to collect target data information, obtain a historical multi-dimensional feature dataset, and construct a substation detection dataset according to the historical multi-dimensional feature dataset; S200: Establish a multi-level feature extraction network, perform multi-object feature extraction on the target data information according to the multi-level feature extraction network, obtain multi-object feature maps, and set up a dilated convolution module to perform feature fusion on the multi-object feature maps to obtain the comprehensive power station features; S300: Build a comprehensive power station detection model based on the multi-level feature extraction network and the dilated convolution module, and train the comprehensive power station detection model according to the power station detection data set; S400: Perform power station detection on the target data information according to the trained comprehensive power station detection model to obtain the power station detection results.

[0023] Specifically, first, remote sensing satellites can be used to obtain the data information of the target area. Remote sensing satellites can acquire the geographical and power infrastructure data of a large area, including the distribution information of substation locations, surrounding environments, buildings, trees and other obstacles. After collecting the target data information, the system integrates these data with the historical data set, and constructs a power station detection data set for subsequent detection according to the historical multi-dimensional feature data set. This data set contains the multi-dimensional features of the substation, such as the type of power facilities, relative position, surrounding environment type, etc. Build a multi-level feature extraction network, which can extract multiple target features in the target data information, including the structural features of the substation, surrounding environment features, etc. This feature extraction network combines deep learning technology and gradually extracts features from low level to high level through a multi-layer convolutional network. Then, by setting up a dilated convolution module to perform feature fusion on the extracted multi-object feature maps, the dilated convolution module can capture target features of different sizes at different sampling intervals, enhancing the adaptability of the network to substations of different scales. This fusion process can effectively integrate the key features of the substation and its surrounding environment, and eliminate the noise interference in the environment. Based on the multi-object feature maps and the feature fusion results of the dilated convolution module, build a comprehensive power station detection model. The goal of this model is to be able to automatically identify the substation in the target data through training, and accurately distinguish the substation from other similar objects such as surrounding buildings and trees. During the training process, the model can continuously adjust parameters to minimize the error function and gradually improve the detection accuracy. Through the trained comprehensive power station detection model, the substation detection can be performed on the target data information, and finally the detection results of the substation can be obtained. These detection results include the specific location of the substation, the description of the surrounding environment, and the relevant power facility information.

[0024] Through the technical solution of the present invention, the problems that the traditional detection method is easily interfered by factors such as weather and terrain in a complex environment, resulting in unstable recognition effect and low accuracy are effectively solved. By adopting the combination of a multi-level feature extraction network and a dilated convolution module, the fusion ability of multi-level features is effectively improved, thus ensuring high-precision recognition of substations and their equipment. At the same time, the training and optimization of the deep learning model enable the present invention to operate stably in a complex environment, significantly improving the accurate positioning of power station facilities and meeting the high-precision recognition requirements.

[0025] Furthermore, as Figure 2 shown, multi-target feature extraction is performed on the target data information according to the multi-level feature extraction network to obtain a multi-target feature map, including: S210: Construct a power station feature database according to the historical multi-dimensional feature dataset, and perform feature extraction on the target data information according to the power station feature database to obtain a number of preliminary feature maps, where the preliminary feature maps include a number of low-level feature maps and a number of high-level feature maps; S220: Based on a number of preliminary feature maps, establish residual skip connection paths, which are used to directly link low-level feature maps and high-level feature maps; S230: Set cross-layer feature transfer weights according to a number of preliminary feature maps, and set skip connection parameters according to the residual skip connection paths, and dynamically adjust the cross-layer feature transfer weights based on the skip connection parameters to obtain a multi-target feature map.

[0026] As a preference of the above embodiment, first, a power station feature database is constructed according to the historical multi-dimensional feature dataset. The database contains remote sensing data information of various power station facilities, and annotates different power equipment, facility layouts and structural information. Subsequently, the constructed power station feature database is used to perform feature extraction on the target data information. The preliminary feature maps are divided into low-level feature maps and high-level feature maps. The low-level feature maps mainly capture detailed information (such as the texture and color of equipment), and the high-level feature maps mainly focus on higher-level spatial relationships (such as the geometric shape and arrangement of equipment). Based on the preliminary feature maps, residual skip connection paths are established, which are used to directly link low-level feature maps and high-level feature maps to solve the loss problem in information transmission. The establishment method of the residual skip connection paths can adopt the following method: In some embodiments, the low-level feature and the high-level feature are combined through an addition operation formula, and the addition formula is: , where represents the low-level feature, represents the high-level feature, Denote the merged feature map; fuse the low-level feature map and the high-level feature map through residual connection, which ensures that the low-level features are directly passed to the deep network without going through multiple convolution and pooling operations. This connection method helps the network maintain attention to low-level details at deeper levels; then, by introducing a weight coefficient, control the contributions of low-level and high-level features during the fusion process. Initially, set a fixed cross-layer transfer weight, and then dynamically adjust these weights according to the training data. During the training process, the weights of cross-layer feature transfer can be dynamically adjusted through the backpropagation algorithm. Weightedly fuse all optimized features to obtain the final multi-objective feature map, which will contain comprehensive information about multiple power facilities (such as substations, transformers, switchgear, etc.) in the remote sensing image.

[0027] Furthermore, as Figure 3 shown, dynamically adjust the cross-layer feature transfer weight based on the skip connection parameters, including: S231: Calculate the feature similarity between the low-level feature map and the high-level feature map according to the power station feature database, and set the initial cross-layer feature transfer weight according to the feature similarity; S232: Calculate the cross-layer influence factor according to the initial cross-layer feature transfer weight, and correct the skip connection parameters according to the cross-layer influence factor; S233: Re-distribute the cross-layer feature transfer weight based on the corrected skip connection parameters.

[0028] In this embodiment, obtain the low-level feature map and the high-level feature map of the target data information through a feature extraction network (such as a convolutional neural network), and then evaluate the spatial and semantic similarity between the two by calculating the similarity between the low-level feature map and the high-level feature map. The following similarity calculation method can be adopted: , Denote the low-level feature, Denote the high-level feature, Denote the dot product operation, Denote the norm of the vector; according to the calculated feature similarity, assign initial cross-layer feature transfer weights to the low-level and high-level features, then calculate the cross-layer influence factor and correct the skip connection parameters. The cross-layer influence factor is used to quantify the relationship strength between the low-level and high-level features. Through the backpropagation algorithm, calculate the gradient of the loss function with respect to the features of each layer of the network. During the backpropagation process, use the chain rule to calculate the gradient layer by layer. The magnitude of the gradient reflects the sensitivity of the feature to the loss, that is, the impact of feature changes on the final output result. If the gradient is large, it means that the feature of this layer has a large impact on the final prediction result; conversely, if the gradient is small, the impact of this layer of feature is small. Based on the calculated gradient magnitude, the cross-layer influence factor can be defined to quantify the relative influence degree of the low-level and high-level features on the final output; correct the skip connection parameters according to these cross-layer influence factors. During the entire training process, the cross-layer influence factor will be continuously adjusted according to the gradient information of the backpropagation. After each training iteration, the model will recalculate the gradients of the low-level and high-level features, and then adjust the cross-layer influence factor and the skip connection parameters.

[0029] Furthermore, perform power station detection on the target data information according to the trained comprehensive power station detection model, including: Extract the multi-level feature maps of the target data information, where the low-level feature maps contain the texture details of the power equipment, and the high-level feature maps contain the geometric contours of the power equipment; Perform feature weighted integration on the low-level feature maps and the high-level feature maps to obtain enhanced feature maps; Perform local area scanning on the enhanced feature maps to determine the equipment area, and extract the physical features and spatial coordinates of the power equipment in the equipment area; Obtain the historical equipment relationships of the power station, perform consistency verification on the power equipment in the equipment area according to the historical equipment relationships of the power station, and output the power station facility location information.

[0030] Specifically, through the obtained target data information, the target recognition algorithm is used to identify facilities in the target data information. An improved Yolov8 network can be used to construct a target recognition model: EfficientNet is used as the backbone network for feature extraction. EfficientNet can extract low-level features in the image through its efficient convolutional layer design, such as the texture and local details of power equipment. Through multiple convolutional layers, local information of power equipment such as surface texture, details, and edges is extracted. In the deep part of the YOLOv8 network, deep convolutional layers and Transformer Blocks are used to enhance the global feature capture ability and extract the geometric contours and spatial relationships of power equipment. Subsequently, BiFPN (Bidirectional Feature Pyramid Network) is used to weight and integrate low-level and high-level features. By combining low-level and high-level features, an enhanced feature map is obtained through a weighted method. Based on the fused enhanced feature map, the target detection module of the YOLOv8 network is used to perform local area scanning to identify the specific location and area of power equipment. Finally, using the historical equipment relationship data of the power station, information such as the spatial location, equipment type, and facility layout between equipment is associated. Based on the historical equipment relationship database, consistency verification is performed on the detected equipment area. The verification process includes verifying information such as the type, location, and size of the equipment to ensure the consistency of the recognition result with historical data. After consistency verification, the correct location and type of power equipment are finally determined, and the power station facility positioning information is output, including the specific coordinates, type, function, etc. of the equipment.

[0031] Furthermore, a comprehensive power station detection model is constructed, including: According to the power station facility positioning information, the comprehensive power station features are matched and compared to obtain the spatial correspondence between the power station facility positioning information and the comprehensive power station features; If the power station facility positioning information does not match the comprehensive power station features, the feature contribution weight is calculated based on the power station facility positioning information, and the category distribution of the comprehensive power station features is corrected according to the feature contribution weight to match the power station facility positioning information. The feature contribution weight is allocated according to the number of pixels in the regional category; If the power station facility positioning information matches the comprehensive power station features, the power station facility positioning information and the comprehensive power station features are fused to construct a comprehensive power station detection model.

[0032] As a preference of the above embodiments, first, the positions of power station facilities (such as substations, transformers, switchgear, etc.) are marked for the collected target data information, and matched with the comprehensive power station features (such as the global features and local features extracted based on deep learning). The method based on feature embedding can be used to project the power station facility location information and power station features into the same embedding space. In this shared space, the similarity of features will be directly measured. When the power station facility location information does not match the comprehensive power station features, the comprehensive power station features need to be further adjusted. The category distribution of the comprehensive power station features can be corrected by the feature contribution weight, and the following method can be adopted: According to the number of pixels in the area where the power station facility is located (i.e., the size or occupied area of the area), and the distribution of different types of power equipment in this area, calculate the contribution weight of each type of feature. According to the feature contribution weight, correct the category distribution of the comprehensive power station features. For each category, the feature contribution weight can be obtained by the ratio of the number of pixels of this category to the total number of pixels. When the power station facility location information matches the comprehensive power station features, the two are fused to construct a complete power station comprehensive detection model. This model combines the physical features, spatial coordinates of the power station facilities and the comprehensive power station features. The feature-level fusion method can also be used to merge the low-level and high-level features through weighted fusion, splicing or addition operations, etc. After the above steps, the fused power station comprehensive feature model is obtained, and the accurate positioning and identification of the power station facilities are completed.

[0033] Furthermore, as Figure 4 shown, obtain the power station detection result and optimize the image of the power station detection result, including: Obtain the misdetection image and the standard image according to the power station detection result, reshape the features of the misdetection image, and check the standard image; Perform multiple image detail optimizations on the misdetection image to obtain several optimized comparison images with different degrees; Compare each of the several optimized comparison images with the misdetection image respectively, and mark the feature difference areas of the misdetection image; Merge the feature difference areas with the power station detection data set to update the power station detection data set.

[0034] In this embodiment, according to the power station detection results, misdetected images and standard images are obtained. The misdetected images are the incorrect identifications that occur during the model recognition process, which may include misidentified equipment areas or incorrect categories. The standard images are the images obtained based on the annotation information of the actual power station facilities (such as the actual positions and categories of power station equipment). The comparison between the misdetected images and the standard images can adopt the image difference analysis method, and use image quality evaluation methods such as the structural similarity index or the mean square error for comparison to find out the misdetected areas. Subsequently, the image quality can be optimized through multiple Gaussian blur and sharpening processes: for the misdetected images, the Gaussian blur standard deviation is selected based on historical experience, the Gaussian kernel is constructed according to the standard deviation, and the weight value of each pixel point is calculated. Subsequently, the Gaussian kernel is convolved with the misdetected image, each pixel point in the image is traversed, and weighted averaging is performed using the Gaussian weights to output the blurred image. After Gaussian blur, sharpening filtering is performed on the blurred image. In at least one embodiment, the Laplace operator is used for the sharpening operation. Through repeated optimization, multiple optimized comparison images are obtained, and these images vary in the degree of detail restoration, and may include comparison results of high-quality optimization, medium-quality optimization, and low-quality optimization: the optimization process can be processed through a hierarchical optimization strategy. Each time optimization is performed, the hyperparameters of the network (such as the learning rate, batch size, etc.) can be adjusted to control the optimization effect. Through multiple optimizations, comparison images at different levels are obtained, and the multiple optimized comparison images are compared with the original misdetected image one by one to mark the characteristic difference areas of the misdetected image. The difference areas refer to the obvious different parts between the optimized image and the misdetected image. These areas usually contain partial information of the misdetection, and more accurate identification can be obtained through optimization. The comparison process can be carried out through pixel-level difference comparison; the characteristic difference areas marked in the misdetected image can be merged with the power station detection data set by using the incremental learning method, and the updated data set will contain more optimized information, especially through learning and improvement in the misidentification, so as to provide more training data for the model.

[0035] Furthermore, the verification of the standard image includes: Collect the historical ledger coordinates, extract the centroid coordinates of the power station area according to the standard image, compare the historical ledger coordinates with the centroid coordinates of the power station area, and obtain the plane distance difference; Set the offset threshold according to the historical equipment relationship of the power station. If the plane distance difference exceeds the offset threshold, perform secondary verification on the standard image to obtain the corrected offset; Correct the standard image according to the corrected offset and verify it again until the plane distance difference meets the offset threshold.

[0036] As an optimization of the above embodiments, first, historical ledger coordinates are collected according to a Geographic Information System (GIS) or manual calibration method. The accurate power station area can be obtained through image segmentation techniques (such as threshold segmentation, edge detection, etc.). Subsequently, the centroid of this area is calculated. The calculation of the centroid can be completed by weighted averaging the coordinates of all pixels in the image. The planar distance difference between the historical ledger coordinates and the centroid coordinates of the power station area is calculated, and the planar distance difference can be calculated using the Euclidean distance formula. Subsequently, an initial offset threshold can be set based on historical experience. Through the verification test of some power stations, check whether the offset threshold is appropriate. If it is found that too many errors are not corrected, the offset threshold may need to be reduced; if the errors are small, the threshold can be appropriately increased; if the planar distance difference exceeds the offset threshold, a secondary verification is performed. During this process, the positioning of the power station area is reexamined to obtain the corrected offset: through image translation operations, the power station area is moved according to the calculated corrected offset so that it matches the historical ledger coordinates, and then verified through specific markers in the image (such as the specific location of the transformer, equipment identification) to ensure that the corrected power station position is more accurate. After obtaining the corrected offset, the standard image is corrected according to the corrected offset. After completion of the correction, the planar distance difference is recalculated and verified again. This process is continuously iterated until the planar distance difference meets the offset threshold.

[0037] Furthermore, the power station comprehensive detection model is trained according to the power station detection data set, including: Clean the historical multi-dimensional feature data set to obtain the original training set; Perform data information processing on several historical multi-dimensional feature data sets respectively according to the original training set to construct the power station detection data set. The data information processing includes target recognition and manual correction; Divide the power station detection data set into a power station training set and a power station test set. Train the power station comprehensive detection model according to the power station training set and verify the power station comprehensive detection model according to the power station test set.

[0038] In this embodiment, before starting model training, it is necessary to clean the historical multi-dimensional feature dataset. For images with poor quality, image enhancement techniques (such as Gaussian blur denoising, contrast enhancement, etc.) can be used to remove noise and improve image quality. At the same time, check whether the annotation of each image is accurate. If annotation errors are found (such as incorrectly annotating the location of power station facilities), manual correction is required. If some images cannot clearly identify power station facilities or the annotation area is too small, they can be removed from the dataset. Then, a target recognition algorithm (Feature Pyramid Network (FPN)) is used to perform target recognition on each image and mark the location of power station facilities (for example, transformers, switchgear, substation buildings, etc.). For power station equipment that is difficult to identify, the diversity of the training set can be increased by augmenting data (such as image flipping, rotation, cropping, etc.) to obtain the final training set. The final training dataset is divided into a dataset according to a ratio of 6:4 to obtain the processed training set and test set, and the power station comprehensive detection model is trained and verified.

[0039] Embodiment 2; Based on the same inventive concept as a satellite image processing system in the foregoing embodiment, the present invention also provides a satellite image processing system, as Figure 5 shown, the system includes: A data information construction module, which uses a remote sensing satellite to collect target data information, obtains a historical multi-dimensional feature dataset, and constructs a power station detection dataset according to the historical multi-dimensional feature dataset; A feature extraction and fusion module, which establishes a multi-level feature extraction network, performs multi-target feature extraction on the target data information according to the multi-level feature extraction network to obtain a multi-target feature map, and sets a dilated convolution module to perform feature fusion on the multi-target feature map to obtain a comprehensive power station feature; An identification model construction and training module, which constructs a power station comprehensive detection model based on the multi-level feature extraction network and the dilated convolution module, and trains the power station comprehensive detection model according to the power station detection dataset; A power station identification and output module, which performs power station detection on the target data information according to the trained power station comprehensive detection model to obtain a power station detection result.

[0040] The above adjustment system in the present invention can effectively implement the intelligent detection method for substations, and the technical effects that can be achieved are as described in the above embodiments, which will not be elaborated here.

[0041] Furthermore, the feature extraction and fusion module includes: A power station feature extraction and analysis unit, which constructs a power station feature database according to the historical multi-dimensional feature dataset, performs feature extraction on the target data information according to the power station feature database to obtain a number of preliminary feature maps, and the preliminary feature maps include a number of low-level feature maps and a number of high-level feature maps; A residual connection feature transfer unit that respectively establishes residual skip connection paths based on a plurality of preliminary feature maps, and the residual skip connection paths are used to directly link low-level feature maps and high-level feature maps; A cross-layer weight dynamic adjustment unit that sets cross-layer feature transfer weights according to a plurality of preliminary feature maps, sets skip parameters according to the residual skip connection paths, and dynamically adjusts the cross-layer feature transfer weights based on the skip parameters to obtain multi-target feature maps.

[0042] Similarly, for the above optimization solutions of the system, the corresponding optimization effects of the methods in Embodiment 1 can also be respectively achieved, and details are not described herein again.

[0043] Although the present application has been described in connection with specific features and their embodiments, it is obvious that various modifications and combinations can be made without departing from the spirit and scope of the present application. Accordingly, the present specification and the drawings are merely exemplary illustrations of the present application as defined by the appended claims, and are considered to cover any and all modifications, variations, combinations, or equivalents within the scope of the present application. Obviously, those skilled in the art can make various changes and modifications to the present application without departing from the scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the present application and its equivalent technologies, the present application is intended to include these changes and modifications.

Claims

1. A substation intelligent detection method, characterized in that: The method comprises: Collect target data information, obtain a historical multidimensional feature data set, and construct a power plant detection data set based on the multidimensional feature data set; Establishing a multi-level feature extraction network, performing multi-target feature extraction on the target data information according to the multi-level feature extraction network, obtaining a multi-target feature map, and setting a hole convolution module to perform feature fusion on the multi-target feature map to obtain comprehensive power station features; Building a comprehensive power station detection model based on the multi-level feature extraction network and the dilated convolution module, and training the comprehensive power station detection model according to the power station detection data set; The power station detection is performed on the target data information according to the trained power station comprehensive detection model to obtain the power station detection result.

2. The intelligent detection method for substation according to claim 1, characterized in that: Performing multi-target feature extraction on the target data information according to the multi-level feature extraction network to obtain a multi-target feature map includes: Building a power plant feature database according to the historical multidimensional feature data set, performing feature extraction on the target data information according to the power plant feature database, and obtaining a number of preliminary feature maps, wherein the preliminary feature maps include a number of low-level feature maps and a number of deep-level feature maps; Based on the plurality of preliminary feature maps, respectively establish residual skip paths, wherein the residual skip paths are used to directly link the low-layer feature map and the deep-layer feature map; A cross-layer feature transfer weight is set according to the plurality of preliminary feature maps, a jump parameter is set according to the residual jump path, and the cross-layer feature transfer weight is dynamically adjusted based on the jump parameter to obtain a multi-target feature map.

3. The intelligent detection method for substation according to claim 2, characterized in that: Dynamically adjusting the cross-layer feature transfer weight based on the skip parameter includes: Calculating feature similarity between the low-layer feature map and the deep-layer feature map according to the power station feature database, and setting initial cross-layer feature transfer weights according to the feature similarity; Calculating a cross-layer influence factor according to the initial cross-layer feature transfer weight, and modifying the jump connection parameter according to the cross-layer influence factor; The cross-layer feature transfer weights are redistributed based on the modified skipping parameters.

4. The intelligent detection method for substation according to claim 1, characterized in that: Performing power station detection on target data information according to the trained power station comprehensive detection model includes: Extracting a multi-level feature map of the target data information, wherein a low-level feature map contains texture details of the power equipment, and a high-level feature map contains a geometric outline of the power equipment; Performing feature weighted integration on the low-level feature map and the high-level feature map to obtain an enhanced feature map; Scanning a local area of ​​the enhanced feature map to determine a device area, and extracting physical features and spatial coordinates of the power equipment in the device area; The historical equipment relationship of the power station is obtained, the consistency check of the power equipment in the equipment area is performed according to the historical equipment relationship of the power station, and the positioning information of the power station facilities is output.

5. The intelligent detection method for substation according to claim 4, characterized in that: Construct a comprehensive power plant detection model, including: According to the power station facility positioning information, matching and comparing the comprehensive power station characteristics are performed to obtain a spatial correspondence between the power station facility positioning information and the comprehensive power station characteristics; If the power station facility location information does not match the comprehensive power station feature, a feature contribution weight is calculated based on the power station facility location information, and the category distribution of the comprehensive power station feature is corrected according to the feature contribution weight to match the power station facility location information, and the feature contribution weight is allocated according to the number of pixels of the regional category; If the power station facility positioning information matches the comprehensive power station characteristics, the power station facility positioning information and the comprehensive power station characteristics are integrated to construct a comprehensive power station detection model.

6. The intelligent detection method for substation according to claim 4, characterized in that: Obtaining power station detection results and performing image optimization on the power station detection results, including: Acquire a false detection image and a standard image according to the power station detection result, reshape the features of the false detection image, and verify the standard image; Performing multiple image detail optimizations on the misdetected image to obtain a number of optimized comparison images of different degrees; Compare the plurality of optimized comparison images with the false detection images respectively, and mark the feature difference areas of the false detection images; The characteristic difference region is merged with the power station detection data set to update the power station detection data set.

7. The intelligent detection method for substation according to claim 6, characterized in that: The standard image is checked, including: Collecting historical record coordinates, and extracting the centroid coordinates of the power station area according to the standard image, comparing the historical record coordinates with the centroid coordinates of the power station area, and obtaining the plane distance difference; An offset threshold is set according to the historical equipment relationship of the power station, and if the plane distance difference exceeds the offset threshold, a secondary verification is performed on the standard image to obtain a corrected offset; The standard image is corrected according to the corrected offset and calibrated again until the plane distance difference meets the offset threshold.

8. The intelligent detection method for substation according to claim 1, characterized in that: The power station comprehensive detection model is trained according to the power station detection data set, including: Cleaning the historical multidimensional feature data set to obtain an original training set; According to the original training set, data information processing is performed on the plurality of historical multi-dimensional feature data sets to construct a power station detection data set, wherein the data information processing includes target recognition and manual correction; The power station detection data set is divided into a power station training set and a power station test set, the power station comprehensive detection model is trained according to the power station training set, and the power station comprehensive detection model is verified according to the power station test set.

9. A satellite image processing system, characterized in that: The system comprises: A data information construction module, which uses remote sensing satellites to collect target data information and obtains a historical multidimensional feature data set, and constructs a power station detection data set based on the historical multidimensional feature data set; A feature extraction and fusion module is used to establish a multi-level feature extraction network, perform multi-target feature extraction on the target data information according to the multi-level feature extraction network, obtain a multi-target feature map, and set a hole convolution module to perform feature fusion on the multi-target feature map to obtain comprehensive power station features; A recognition model construction training module is used to construct a comprehensive power station detection model based on the multi-level feature extraction network and the dilated convolution module, and to train the comprehensive power station detection model according to the power station detection data set; The power station identification and output module performs power station detection on the target data information according to the trained power station comprehensive detection model to obtain the power station detection result.

10. The satellite image processing system according to claim 9, characterized in that: The feature extraction and fusion module comprises: A power plant feature extraction and analysis unit, which constructs a power plant feature database according to the historical multidimensional feature data set, extracts features of the target data information according to the power plant feature database, and obtains a number of preliminary feature maps, wherein the preliminary feature maps include a number of low-level feature maps and a number of deep-level feature maps; A residual connection feature transfer unit, which establishes residual skip paths based on the plurality of preliminary feature maps, respectively, wherein the residual skip paths are used to directly link the low-level feature maps with the deep-level feature maps; A cross-layer weight dynamic adjustment unit sets a cross-layer feature transfer weight according to the plurality of preliminary feature maps, sets a jump parameter according to the residual jump path, and dynamically adjusts the cross-layer feature transfer weight based on the jump parameter to obtain a multi-target feature map.

Images