Power grid hidden danger identification method and system based on deep learning

Abstract

The invention provides a power grid hidden danger identification method and system based on deep learning, and belongs to the field of power grid safety monitoring. Multi-type hidden danger recognition is realized based on the pre-trained power grid hidden danger recognition model, and the recognition accuracy, the response time and the missing recognition rate are balanced through dynamic optimization during model pre-training. The method comprises the steps of firstly obtaining power grid multi-dimensional image data covering space-air-ground cross-space hierarchy, then carrying out standardized preprocessing on the data, then extracting local detail features and cross-region global association features in a region, carrying out fusion to form multi-dimensional fusion feature data, finally inputting the multi-dimensional fusion feature data into a trained model, and outputting a hidden danger identification result. Through multi-source image feature deep fusion and model dynamic optimization, effective integration of multi-source data and accurate hidden danger recognition are realized, the accuracy and timeliness of power grid hidden danger recognition are remarkably improved, and the actual requirements of power grid operation and maintenance are fully met.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of power grid safety monitoring, and particularly relates to a power grid hidden danger identification method and system based on deep learning. BACKGROUND

[0002] The power grid plays a core role in energy transmission in the national economy and social operation, and its safe and stable operation has a direct and key impact on industrial production continuity, livelihood electricity guarantee and public safety order. With the continuous development of science and technology, the technology for guaranteeing the safe operation of the power grid is also continuously improving.

[0003] However, the environment of the current power grid coverage area is extremely complex, there are many uncertain factors, and there are also many human disturbances. In such a situation, accurate identification of power grid hidden dangers becomes a key link to guarantee the safe operation of the power grid. The existing power grid hidden danger identification technology faces many challenges when dealing with complex environments and disturbances.

[0004] Most importantly, the multi-source monitoring data in the existing technology are in a fragmented state, lack effective standardization fusion mechanisms, and are excessively dependent on traditional algorithms. This makes the power grid hidden danger identification perform poorly in accuracy, response speed and missed identification rate, and it is difficult to meet the requirements of the power grid for high-precision and fast-response hidden danger identification. SUMMARY

[0005] Therefore, the application provides a power grid hidden danger identification method and system based on deep learning, aiming to solve the multi-source monitoring data fusion problem, improve the accuracy and timeliness of power grid hidden danger identification, and meet the operation and maintenance needs of the power grid.

[0006] In order to achieve the above-mentioned purpose, the technical scheme provided by the application is as follows:

[0007] In a first aspect, the application provides a power grid hidden danger identification method based on deep learning, which is realized based on a pre-trained power grid hidden danger identification model. The power grid hidden danger identification model is used to identify different types of power grid hidden dangers, and in the pre-training process, the power grid hidden danger identification model is dynamically optimized according to the identification effect, so that the model meets the balance requirements among identification accuracy, response time and missed identification rate.

[0008] The method comprises:

[0009] Obtaining power grid space-time-land multi-dimensional image data; the space-time-land multi-dimensional image data is cross-space level data covering space, time and land;

[0010] The air-space-ground multi-dimensional image data is standardized and preprocessed to obtain standardized data, and local detail feature extraction within a region and global correlation feature extraction across regions are performed on the standardized data, and multi-dimensional fusion feature data is obtained by fusing the extracted features.

[0011] The multi-dimensional fusion feature data is input into the trained power grid hidden danger identification model to obtain an identification result.

[0012] Further, the air-space-ground multi-dimensional image data at least includes:

[0013] Satellite images in the space domain, unmanned aerial vehicle images in the air domain, and ground infrared sensor images and monitoring images in the region.

[0014] Further, the air-space-ground multi-dimensional image data is standardized and preprocessed to obtain standardized data, including:

[0015] The format and pixel arrangement rules of the air-space-ground multi-dimensional image data are unified to obtain first image data;

[0016] The image data of different resolutions in the first image data is scaled matched by a bilinear interpolation algorithm to obtain second image data;

[0017] The second image data is subjected to radiation correction, geometric correction and fusion processing to obtain standardized data.

[0018] Further, the standardized data is subjected to local detail feature extraction within a region and global correlation feature extraction across regions, including:

[0019] Local detail feature extraction within a region is performed by a convolutional neural network, and the extracted local detail features at least include tower inclination angle, conductor hanging object profile, ground water area and equipment abnormal heating area;

[0020] Global correlation feature extraction across regions is performed by a Transformer architecture, and the extracted global correlation features at least include the spatial distance between buildings and power lines, cross-region terrain change correlation and disaster scene spread trend.

[0021] Further, in the pre-training process, the power grid hidden danger identification model is dynamically optimized according to the identification effect to meet the balance requirements among identification accuracy, response time and missed identification rate, including:

[0022] Setting a normalized benchmark value for response time and missed identification rate;

[0023] Based on the normalized benchmark value, the actual data of response time and missed identification rate generated by model identification is normalized;

[0024] The model recognition accuracy, the normalized response time and the missed recognition rate are weighted according to preset weight coefficients to obtain a comprehensive performance result.

[0025] The model is continuously adjusted and trained until the comprehensive performance meets the standard or the iteration number reaches the upper limit, so as to complete dynamic optimization.

[0026] In a second aspect, the present application provides a power grid hidden danger identification system based on deep learning, which is realized based on a pre-trained power grid hidden danger identification model.

[0027] The system comprises:

[0028] A multi-source image acquisition unit is configured to acquire power grid space-time-ground multi-dimensional image data.

[0029] A deep learning processing unit is configured to perform standardization preprocessing on the space-time-ground multi-dimensional image data to obtain standardized data, and perform local detail feature extraction within a region and global correlation feature extraction across regions on the standardized data, and fuse the extracted features to obtain multi-dimensional fusion feature data.

[0030] A power grid hidden danger identification unit is configured to input the multi-dimensional fusion feature data into the trained power grid hidden danger identification model to obtain a hidden danger identification result.

[0031] Further, in the multi-source image acquisition unit, the space-time-ground multi-dimensional image data at least comprises:

[0032] Satellite images of the sky, unmanned aerial vehicle images of the space, and ground infrared sensor images and monitoring images of the region.

[0033] Further, in the deep learning processing unit, the standardization preprocessing on the space-time-ground multi-dimensional image data to obtain the standardized data comprises:

[0034] Uniformly processing the space-time-ground multi-dimensional image data in format and pixel arrangement rules to obtain first image data;

[0035] Performing scale matching on image data of different resolutions in the first image data by a bilinear interpolation algorithm to obtain second image data;

[0036] Performing radiation correction, geometric correction and fusion processing on the second image data to obtain the standardized data.

[0037] Further, in the deep learning processing unit, the standardized data is subjected to intra-regional local detail feature extraction and cross-regional global correlation feature extraction, including:

[0038] The local detail feature extraction is performed by a convolutional neural network, and the extracted local detail features at least include a tower inclination angle, a conductor hanging profile, a ground water range, and an equipment abnormal heating area;

[0039] The global correlation feature extraction is performed by a Transformer architecture, and the extracted global correlation features at least include a spatial distance between a building and a power grid line, a cross-regional terrain change correlation, and a disaster scene spreading trend.

[0040] Further, in the pre-training process, the power grid hidden danger identification model is dynamically optimized according to the identification effect, so that the model meets the balance requirements among the identification accuracy, response time and missed identification rate, including:

[0041] Setting a normalized benchmark value for the response time and the missed identification rate;

[0042] Based on the normalized benchmark value, the actual data of the response time and the missed identification rate generated by the model identification is normalized;

[0043] According to a preset weight coefficient, the identification accuracy of the model is weighted with the normalized response time and missed identification rate to obtain a comprehensive performance result;

[0044] Taking the optimal comprehensive performance result as the target, the model is continuously adjusted and trained until the comprehensive performance meets the standard or the iteration number reaches the upper limit, so as to complete the dynamic optimization.

[0045] In summary, the application provides a power grid hidden danger identification method and system based on deep learning, which is realized based on a pre-trained power grid hidden danger identification model. BRIEF DESCRIPTION OF DRAWINGS

[0046] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, a brief introduction will be given to the drawings needed in the embodiments or prior art description. Obviously, the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor on the basis of these drawings.

[0047] Figure 1 A flowchart of a power grid hidden danger identification method based on deep learning provided by the embodiment of the present application;

[0048] Figure 2 A flowchart of a power grid hidden danger identification system based on deep learning provided by the embodiment of the present application;

[0049] Figure 3 A composition block diagram of a computer device provided by the embodiment of the present application. DETAILED DESCRIPTION

[0050] In order to make the purposes, characteristics and advantages of the present application more obvious and easy to understand, the technical solutions in the embodiments of the present application will be described clearly and completely below in combination with the drawings in the embodiments of the present application. Obviously, the embodiments described below are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of the present application.

[0051] Please refer to Figure 1 The embodiment provides a power grid hidden danger identification method based on deep learning. The method is realized based on a pre-trained power grid hidden danger identification model. The power grid hidden danger identification model is used to identify different types of power grid hidden dangers. In the pre-training process, the power grid hidden danger identification model is dynamically optimized according to the identification effect, so that the model meets the balance requirements among the identification accuracy, response time and missed identification rate.

[0052] It can be understood that the model introduces a dynamic optimization mechanism in the pre-training process, and does not use a fixed parameter training mode. Instead, the identification effect is used as real-time feedback, the performance relationship among the identification accuracy, response time and missed identification rate is balanced in a targeted manner by adjusting the internal parameters of the model (such as feature matching weight, inference calculation parameter and the like), so as to ensure that the model neither sacrifices the response speed for the sake of accuracy, nor leads to a too high missed identification rate for the sake of fast response. At the same time, the pre-training process of the model is also adapted to the cross-space level characteristics of space-time-ground multi-dimensional image data, and the details and global attributes of multi-dimensional fusion features, so as to capture the feature rules of different types of power grid hidden dangers.

[0053] The method comprises the following steps.

[0054] S11: Obtain power grid space-time-ground multi-dimensional image data. The space-time-ground multi-dimensional image data is cross-space level data covering the space domain, the sky domain and the region.

[0055] It should be noted that the space-time-ground multi-dimensional image data refers to a multi-source image set covering the space domain, the sky domain and the region. The sky domain data is usually satellite image (which can cover a large area), the space domain data is unmanned aerial vehicle image (which can cover the power grid corridor at a medium or close distance), and the region data is ground sensor / monitoring image (which can cover the details of the equipment at a close distance).

[0056] S12: Standardize the space-time-ground multi-dimensional image data to obtain standardized data, and extract local detail features in the region and global correlation features across regions from the standardized data, and fuse the extracted features to obtain multi-dimensional fusion feature data.

[0057] It should be noted that the standardized preprocessing refers to the process of unified processing of multi-source heterogeneous space-ground-earth image data, aiming to eliminate the heterogeneity of different data sources in terms of format, scale, accuracy, etc., so that the data meet the input requirements of subsequent feature extraction.

[0058] Local detail feature extraction refers to extracting micro features (such as tower inclination angle, conductor hanging profile, etc.) that reflect the specific form of hidden dangers and the state of equipment for power grid equipment and surrounding small range areas (such as single tower and local line section).

[0059] Global correlation feature extraction refers to extracting macro features (such as the spatial distance between buildings and lines, and the correlation of cross-regional terrain changes) that reflect the spatial relationship and influence trend between different regions for cross-regional and large-scale power grid scenes.

[0060] Multi-dimensional fusion feature data is a comprehensive feature set formed by integrating local detail features and global correlation features, which has both micro detail recognition and macro correlation logic.

[0061] S13: input the multi-dimensional fusion feature data into the trained power grid hidden danger identification model to obtain the hidden danger identification result.

[0062] It should be noted that the power grid hidden danger identification model refers to a deep learning model (such as CNN+Transformer fusion architecture) that has been pre-trained and dynamically optimized, which has learned the feature rules of various power grid hidden dangers.

[0063] The embodiment provides a power grid hidden danger identification method based on deep learning. The method uses the feature learning and reasoning ability of the deep learning model to convert the preprocessed fusion features into identification results. The trained power grid hidden danger identification model has stored feature templates of various hidden dangers through the pre-training process. When the multi-dimensional fusion feature data is input, the model will compare the similarity of the fusion features and the hidden danger feature templates through feature matching, probability calculation and other reasoning processes, and finally output the corresponding hidden danger identification result. At the same time, since the model has balanced the identification accuracy, response time and missed identification rate through dynamic optimization during the pre-training stage, it can quickly respond while reducing false positives and missed cases, ensuring the practicality and reliability of the identification results. The method uses a dual feature extraction and fusion strategy of local detail features within the region and global correlation features across regions, which captures specific features of equipment micro defects and excavates correlation rules of cross-regional scenes, achieving comprehensive coverage of feature information. Moreover, the model introduces a dynamic optimization mechanism during the pre-training process, which adjusts the model parameters in real time according to the identification effect, ensuring that the three are optimally balanced, improving the accuracy and timeliness of power grid hidden danger identification, and meeting the needs of power grid operation and maintenance for high-quality hidden danger identification.

[0064] In an embodiment of the present application, the space-time-ground multi-dimensional image data at least includes:

[0065] Satellite images of the space domain, unmanned aerial vehicle images of the air domain, and ground infrared sensor images and monitoring images of the land domain.

[0066] Illustratively, the space-time-ground multi-dimensional image data can include satellite images accessing high-resolution remote sensing satellite and SAR satellite data, unmanned aerial vehicle images linked with power grid nests, and ground infrared sensor and video monitoring data.

[0067] When acquiring satellite images, a data interface is first established with a service provider to access high-resolution remote sensing satellite data such as Gaofen 6 and SAR satellite data such as Gaofen 3. For example, in a typhoon disaster prevention scenario in a certain area, Gaofen 6 satellite takes pictures according to the preset disaster survey area coordinates to obtain optical images with a resolution of 0.8 m, and SAR satellite penetrates through clouds in cloudy weather to synchronously obtain radar images with a resolution of 1 m in the area. Specifically, Gaofen 6 satellite is selected to take pictures of the area covered by the transmission line, and the image resolution parameter is set to 0.8 m. At this resolution, each pixel corresponds to an actual area of 0.8 m x 0.8 m on the ground. After taking pictures, the image data is received by the satellite data receiving station, and after format conversion, it is stored in the system database to ensure that the resolution is ≤1 m. Gaofen 3 satellite is selected for SAR (if in cloudy weather caused by a typhoon, the cloud-penetrating observation mode is enabled, in which the satellite transmits microwave signals and receives ground reflected signals, and is not affected by cloud cover), and the surrounding area of the substation is photographed to obtain SAR images with a resolution of 1 m, clearly showing the position and structure outline of the towers in the area.

[0068] When acquiring unmanned aerial vehicle images, the unmanned aerial vehicle is linked with the nest near the substation. The nest pre-stores the coordinate list of the 220 kV transmission line towers in the area, generates a patrol route and sends it to the unmanned aerial vehicle. The unmanned aerial vehicle carries a 20 million pixel camera, flies along the route at a speed of 8 m / s and a height of 100 m, and acquires high-definition tower images with a resolution of 0.08 m. Specifically, the nest pre-stores detailed route data of the towers on the line, including tower numbers (T1001-T1010), coordinates of each tower (such as T1001: 113.5°E, 22.3°N; T1002: 113.52°E, 22.31°N), flight height (100 m), flight speed (8 m / s), etc. After the unmanned aerial vehicle takes off from the nest, it receives the route data sent by the nest and flies along the preset route. The 20 million pixel industrial camera carried by the unmanned aerial vehicle is set to continuous shooting mode, and a tower image is taken every 50 meters. By calculating the proportional relationship between the image pixels and the actual ground distance, the image resolution is obtained as 0.08 m, which meets the requirement of ≥0.1 m. For example, in the T1005 tower image taken, the bolt structure with a diameter of 5 cm on the tower cross arm can be clearly distinguished, realizing high-definition tower image acquisition.

[0069] When acquiring ground images, an infrared sensor (monitoring range -40℃ 120℃) is installed at the bottom of the tower to collect tower temperature data in real time, and a 4K video monitoring camera is installed at the middle of the tower to transmit one frame of picture every 30 seconds, and the two types of data are integrated to form a ground monitoring data set. Specifically,

[0070] In an embodiment of the present application, the space-time-ground multi-dimensional image data is standardized and preprocessed to obtain standardized data, including:

[0071] S21: The format and pixel arrangement rules of the space-time-ground multi-dimensional image data are unified to obtain first image data.

[0072] Exemplarily, for SAR images (resolution 1m) and optical images (resolution 0.8m) of the same area, first, the file format and pixel arrangement rules of the two are unified.

[0073] S22: The scale of the image data of different resolutions in the first image data is matched by a bilinear interpolation algorithm to obtain second image data.

[0074] Exemplarily, the resolution of the optical image is adjusted to 1m by the bilinear interpolation method to make the pixel sizes of the two types of images consistent, and then the two types of images are subjected to gray scale normalization processing to map the pixel values uniformly to the range of 0-255.

[0075] S23: The second image data is subjected to radiation correction, geometric correction and fusion processing to obtain standardized data.

[0076] Exemplarily, first, radiation correction is performed, and the radiation calibration coefficient K=0.85 and the offset B=-10 of the Gaofen 6 satellite are called to calculate the ground reflectivity by the formula (DN is the original gray value of the image, taking 0-255) For example, when DN=200, L=0.85x200-10=160. Then, known coordinate control points (such as X1=113.5°E, Y1=22.3°N; X2=113.6°E, Y2=22.4°N; X3=113.7°E, Y3=22.3°N) in the region are selected to perform geometric correction and registration of the image. The fusion processing adopts an attention mechanism to calculate the saliency value S of each pixel of each image. The calculation of S uses the formula , wherein is the backscattering coefficient of the SAR image pixel. In the cloud and fog covered area, due to the strong penetration of the SAR image, 0.7, in the cloudless sunny area, 0.3; is the brightness value of the optical image pixel. In the sunny area, the optical image details are clear, Value 0.8, cloud and fog area optical image blur, Value 0.2, according to the calculated S value distribution fusion weight, when S is greater than or equal to 0.5, it is determined that the pixel area is greatly affected by cloud and fog, the SAR image weight is set to 0.6, and the optical image weight is set to 0.4; when S is less than 0.5, it is determined that the pixel area has good lighting conditions, the SAR image weight is set to 0.4, and the optical image weight is set to 0.6; for example, a certain pixel S = 0.62, the final gray value of the pixel after fusion = 0.6 * SAR pixel gray value + 0.4 * optical pixel gray value, and the fusion of SAR and optical images is completed.

[0077] In an embodiment of the present application, CNN+Transformer architecture is used for feature extraction. Local detail feature extraction within the region and global correlation feature extraction across the region are performed on the standardized data, including:

[0078] (1) Local detail feature extraction within the region is performed by a convolutional neural network, and the extracted local detail features at least include tower tilt angle, conductor hanging object profile, ground water area and equipment abnormal heating area.

[0079] For example, the local detail feature extraction is realized by using a CNN layer with a 3*3 convolution kernel (step 1), two convolution layers are set, the first convolution kernel size is 3*3, the step is 1, the number of convolution kernels is 32, and the fused image is subjected to convolution operation; for example, for a pixel (gray value 200) at the edge of a tower in the image, the gray values of the surrounding 8 adjacent pixels are all 100, and the convolution kernel [-1, -1, -1; -1, 8, -1; -1, -1, -1] is multiplied element by element and summed, the calculation process is (100 * (-1)) * 8 + 200 * 8 = -800 + 1600 = 800, the convolution result of the pixel is obtained, the tower edge feature is highlighted, and local feature extraction is realized.

[0080] (2) Global correlation feature extraction across the region is performed by a Transformer architecture, and the extracted global correlation features at least include the spatial distance between buildings and power grid lines, the correlation of cross-regional terrain changes, and the disaster scene spreading trend.

[0081] For example, the global correlation feature extraction is realized by using a Transformer, and the Transformer layer captures global correlation by setting the number of attention heads to 4, converting the feature map output by the CNN layer into a feature vector with a dimension of 64, and each feature vector includes a query vector Q and a key vector K. The attention score of any two pixel feature vectors is calculated as , wherein is the dimension of the feature vector, and the value is 64, , for example The time score = 48 / 8 = 6, and the higher the score, the higher the correlation between the two pixels, and the global correlation between the pixels in different regions of the image is captured in this way.

[0082] In an embodiment of the present application, in the pre-training process, the power grid hidden danger identification model is dynamically optimized according to the identification effect, so that the model meets the balance requirements among identification accuracy, response time and missed identification rate, including:

[0083] S31: Set the normalized reference value of the response time and the missed identification rate.

[0084] The normalized reference value refers to the unified reference standard set for the response time and the missed identification rate (for example, the response time reference value is set to 30 seconds, and the missed identification rate reference value is set to 5%).

[0085] S32: Based on the normalized reference value, the actual data of the response time and the missed identification rate generated by the model identification is normalized.

[0086] The normalization processing refers to taking the set reference value as a reference to convert the response time and the missed identification rate data actually output by the model into dimensionless values in the range of 0-1.

[0087] S33: According to the pre-set weight coefficient, the identification accuracy of the model is weighted with the normalized response time and missed identification rate to obtain the comprehensive performance result.

[0088] For example, the comprehensive performance function is set as , , , , ). F is the comprehensive performance result, P is the identification accuracy (value 0-1), T is the normalized value of the response time (normalized according to T max =30s, value 0-1), R is the missed identification rate (normalized according to R max =5%, value 0-1), for example, P=90% (0.9), T=5s (1-5 / 30≈0.83), R=1% (1-1 / 5≈0.8), then .

[0089] S34: Take the optimal comprehensive performance result as the target, continuously adjust the model and train it until the comprehensive performance meets the standard or the iteration number reaches the upper limit, to complete the dynamic optimization.

[0090] The goal of model training is to achieve the best comprehensive performance result rather than the best single indicator, ensuring that the model achieves a balance between accuracy, response speed, and missed identification control. Model adjustment includes feature matching weights (such as increasing the weight of local detail features to improve accuracy and increasing the weight of global correlation features to reduce missed identification rate), inference calculation parameters (such as learning rate and batch size), and the adjustment direction needs to be optimized according to the short board of the comprehensive performance result (such as if the low comprehensive performance is due to the low missed identification rate normalization value, then focus on improving the global correlation feature weight).

[0091] After the above dynamic optimization, the power grid hidden danger identification model in this embodiment meets the preset requirements for each performance indicator of the five types of hidden danger identification. The specific verification results are as follows:

[0092] First, the requirements for the power grid hidden danger identification model for the five types of hidden danger identification are: the identification accuracy of illegal buildings is ≥91%, the identification accuracy of wind disaster tower collapse, geological disasters, and easy-to-hang objects is ≥86%, the water immersion monitoring accuracy is ≤1m, the identification response time is ≤30s per image, and the missed identification rate is ≤5%.

[0093] For illegal building identification, 100 known illegal building samples (including iron sheet houses, unapproved civilian houses, etc.) around a certain power transmission line corridor are selected. The model successfully identifies 91 illegal buildings. According to the formula "identification accuracy = number of correct identifications / total number of samples", the identification accuracy is 91 / 100 = 91%, which meets the preset requirement of ≥91%.

[0094] For wind disaster tower collapse identification, 80 post-tsunami tower collapse samples (including different inclination angles of 15°-90°) in a certain area are selected. The model successfully identifies 69 tower collapse samples, with an identification accuracy of 69 / 80 = 86.25%, which meets the preset requirement of ≥86%.

[0095] For geological disaster identification, 75 geological disaster samples such as landslides and mudslides in a certain mountainous area are selected. The model identifies 65 valid samples, with an identification accuracy of 65 / 75 ≈ 86.67%, which meets the preset requirement of ≥86%.

[0096] For easy-to-hang object identification, 90 easy-to-hang object samples such as plastic cloth and kites around the line are selected. The model identifies 78 samples, with an identification accuracy of 78 / 90 = 86.67%, which meets the preset requirement of ≥86%.

[0097] For water immersion monitoring accuracy, a number of water immersion monitoring points are selected, the boundary coordinates of the actual water immersion range of each monitoring point are obtained, and the model-identified water immersion boundary is compared with the actual boundary. The maximum distance difference between the two is 0.8m, which meets the monitoring accuracy requirement of ≤1m.

[0098] For the identification response time, 10 power grid images with a resolution of 1920x1080 are selected, and the model start processing time and completion identification time are recorded respectively. The longest processing time is 28 seconds, the shortest is 22 seconds, and the average processing time is 25 seconds, which meets the preset requirement of ≤30s / image. For the missed identification rate, 100 mixed samples covering five types of hidden dangers are selected. After the model processing, the number of missed identification samples is 5. According to the formula "missed identification rate = missed identification number / total sample number", the missed identification rate is 5 / 100 = 5%, which meets the preset requirement of ≤5%.

[0099] In an embodiment of the present application, after generating the hidden danger identification result, a safety distance threshold constraint determination step needs to be performed to determine whether the hidden danger threatens the safety of the power grid, and the safety distance determination deviation of the determination step is ≤0.5m.

[0100] The specific determination process is as follows: first, a preset power grid safety distance threshold database is called, and the threshold of the database is set according to DL / T5092-2013 "110kV-750kV overhead transmission line design specification" (for example, the safety distance threshold of 220kV transmission line and building is 5m, and the safety distance threshold of tree is 3m); then the actual distance between the identified hidden danger and the power grid line is calculated by using an image measurement tool. The calculation method is to select the two closest points of the hidden danger edge and the line in the image, obtain the latitude and longitude coordinates of the two points, and solve the actual distance by using the latitude and longitude distance calculation formula (1° latitude ≈111km, 1° longitude ≈111×cosθ km, θ is the latitude value); finally, the actual distance calculated is compared with the safety distance threshold corresponding to the voltage level, and if the actual distance is less than the threshold, it is determined that the hidden danger threatens the safety of the power grid. For example, a building is identified around the transmission line with a voltage level of 220kV, and the safety distance threshold of 5m is called. The actual distance between the building and the line is calculated to be 4.95m. Since 4.95m<5m, it is determined that the building is a threatening hidden danger. To verify the determination accuracy, 10 samples with known actual safety distance are selected for testing, and the maximum determination deviation is 0.5m, which meets the preset requirement of ≤0.5m.

[0101] In an embodiment of the present application, after the hidden danger identification result is generated, a hidden danger identification and emergency dispatch linkage step needs to be performed to realize the closed-loop processing from hidden danger identification to emergency response, which specifically includes the following process:

[0102] Satellite emergency calling process: when a disaster (such as a typhoon) occurs and large-scale image data needs to be quickly obtained, a calling request containing the coordinates of the disaster investigation area is sent to the satellite service provider to ensure that satellite shooting and data transmission are completed within 6 hours. For example, after the disaster occurs, the service provider arranges satellite shooting within 4 hours and transmits image data to the processing end within 6 hours to provide data support for hidden danger identification in the disaster area.

[0103] Suspected hidden danger unmanned aircraft secondary confirmation process: for suspected hidden danger identified by satellite, start unmanned aircraft verification process, ensure that the total response time from receiving suspected hidden danger information to output verification result ≤1 hour. Specifically: after receiving the suspected hidden danger coordinates, query the unmanned aircraft landing point within 5 km range around the coordinates, dispatch the nearest unmanned aircraft, fly to the hidden danger point at the preset flight speed (such as 10 m / s), flight height (such as 120 m), and shooting resolution (≥0.1 m), complete shooting, and transmit data through the communication network and analyze and verify, for example, after receiving the suspected tower falling coordinates (113.6°E, 22.3°N), the unmanned aircraft flies to the scene for 5 minutes and 20 seconds, and transmits the data after 15 minutes of shooting, the total response time is 28 minutes and 20 seconds, which meets the requirement of ≤1 hour;

[0104] Emergency visualization situation map generation process: collect hidden danger identification results (type, location, level, identification time), power grid equipment related data (tower number, model, operation time, operation and maintenance person in charge and contact method), real-time traffic data (road name, traffic state, nearest parking lot location and remaining parking spaces within 3 km range around the hidden danger point), after standardizing the format of the above data, superimposed on the corresponding regional electronic map layer, an emergency visualization situation map containing multi-dimensional information is generated, which provides intuitive support for dispatching decision.

[0105] In an embodiment of the present application, to realize continuous optimization of the identification model, a model self-upgrade step is also needed, which specifically includes the following processes:

[0106] Sample automatic labeling process: through the mobile terminal tool of the field inspection personnel, collect suspected hidden danger labeling information and image data, the labeling information includes hidden danger type, coordinates, size, distance from tower, etc., after the mobile terminal tool uploads the data to the processing end, the processing end checks the data format and integrity, and after the checking is passed, the sample is added to the labeling sample library, and the labeling personnel, labeling time and other metadata are recorded, ensuring that the number of newly added samples per month is not less than 500, forming a continuously updated labeling sample library;

[0107] Cross-scene image domain alignment processing flow: select a preset number (such as 200) of images of different data sources (satellite, unmanned aircraft), divide them into multiple batches according to the shooting area, calculate the gray mean μ and standard deviation σ of the images of different data sources in each batch, and adjust the gray parameters of one type of image, and then recalculate μ and σ, until the μ difference of the two types of images ≤5, and the σ difference ≤3, eliminating the scene difference of images of different data sources;

[0108] ​Model regular incremental fine-tuning process: set a fixed date to start incremental training in the last month of each quarter, select the preset number (such as 1200) of newly added samples in the sample library of the current quarter, divide them into training set, validation set and test set according to the ratio of 7:2:1, set the training parameters (batch size 32, initial learning rate 0.001, learning rate decay coefficient 0.9, iteration number 60 times), evaluate the model performance every 10 iterations during the training process using the validation set, stop training and save the current optimal model parameters when the validation set accuracy does not improve for 3 consecutive iterations, replace the original model with the new model to realize dynamic optimization of the model, for example, after the training of a certain quarter is completed, the recognition accuracy of the model on the test set is improved by 2.3% compared with the previous update.

[0109] In an embodiment of the present application, the following data transmission and synchronization method is used in the data interaction process to ensure the real-time, integrity and bandwidth adaptability of data transmission:

[0110] Communication protocol and transmission delay control: 5G communication protocol is used for data transmission, multiple test points (such as 10) are selected along the transmission line for transmission delay test, multiple test data packets (such as 100 groups, each group 1KB) are sent at each test point, the time difference between sending and receiving each data packet is recorded to ensure that the average transmission delay is ≤50ms, the test results show that the average delay is 42ms and the maximum delay is 48ms, meeting the real-time transmission requirement;

[0111] External system data synchronization: configure the communication interface parameters of the power grid EMS system and the PMS system, among which the IEC61850 standard protocol is used for the EMS system, the data synchronization period is set to 10 minutes, and the real-time operation data such as tower current, voltage and power are synchronized and transmitted; WebService protocol is used for the PMS system, data synchronization period is set to 1 hour, and data such as equipment account, defect record and maintenance plan are synchronized and transmitted, realizing two-way data synchronization with external power grid system;

[0112] Data compression processing: for non-critical image data (such as daily inspection images without hidden dangers), H.265 algorithm is used for compression, multiple preset resolution (such as 1920x1080) images without hidden dangers are selected for testing, the average size of each image before compression and the average size after compression are counted to ensure that the compression ratio is ≥15:1, the test results show that the average size before compression is 100MB and the average size after compression is 6.4MB, the compression ratio is 15.625:1, effectively reducing the data transmission amount;

[0113] Low bandwidth adaptation strategy: when the transmission bandwidth is less than or equal to 1 Mbps (such as 0.8 Mbps), the key data (single data size 10 KB) such as hidden danger coordinates and types is preferentially transmitted, so as to ensure that multiple key data (such as 8 pieces) can be synchronously transmitted per second, and to avoid delay in transmission of key information in a low bandwidth scene.

[0114] Based on the same inventive concept, the embodiment of the present application also provides a deep learning-based power grid hidden danger identification system for implementing the above-mentioned deep learning-based power grid hidden danger identification method. The implementation scheme for solving the problem provided by the system is similar to the implementation scheme described in the above method, so the specific limitations in the following deep learning-based power grid hidden danger identification system embodiment can be referred to the limitations of the deep learning-based power grid hidden danger identification method in the above text, which will not be repeated here.

[0115] Please refer to Figure 2 The embodiment of the present application also provides a deep learning-based power grid hidden danger identification system, which is implemented based on a pre-trained power grid hidden danger identification model. The power grid hidden danger identification model is used to identify different types of power grid hidden dangers, and in the pre-training process, the power grid hidden danger identification model is dynamically optimized according to the identification effect, so that the model meets the balance requirements among identification accuracy, response time and missed identification rate.

[0116] The system comprises:

[0117] The multi-source image acquisition unit is configured to acquire air-space-ground multi-dimensional image data of the power grid. The air-space-ground multi-dimensional image data is cross-space hierarchical data covering airspace, space and region.

[0118] The deep learning processing unit is configured to perform standardization preprocessing on the air-space-ground multi-dimensional image data to obtain standardized data, and perform local detail feature extraction within a region and global correlation feature extraction across regions on the standardized data, and fuse the extracted features to obtain multi-dimensional fusion feature data.

[0119] The power grid hidden danger identification unit is configured to input the multi-dimensional fusion feature data into the trained power grid hidden danger identification model to obtain a hidden danger identification result.

[0120] Further, in the multi-source image acquisition unit, the air-space-ground multi-dimensional image data at least comprises:

[0121] satellite image of space, unmanned aerial vehicle image of airspace, and ground infrared sensor image and monitoring image of region.

[0122] Further, in the deep learning processing unit, the air-space-ground multi-dimensional image data is standardized and preprocessed to obtain standardized data, including:

[0123] The air, space and ground multi-dimensional image data are unified in format and pixel arrangement rules to obtain first image data;

[0124] The image data of different resolutions in the first image data are matched in scale by a bilinear interpolation algorithm to obtain second image data;

[0125] The second image data are subjected to radiation correction, geometric correction and fusion processing to obtain standardized data.

[0126] Further, in the deep learning processing unit, the standardized data are subjected to local detail feature extraction within a region and global correlation feature extraction across regions, including:

[0127] The local detail feature extraction within a region is performed by a convolutional neural network, and the extracted local detail features at least include a tower inclination angle, a conductor hanging object contour, a ground water area range and an equipment abnormal heating area;

[0128] The global correlation feature extraction across regions is performed by a Transformer architecture, and the extracted global correlation features at least include a spatial distance between a building and a power grid line, a cross-region terrain change correlation and a disaster scene spreading trend.

[0129] Further, in the pre-training process, the power grid hidden danger identification model is dynamically optimized according to the identification effect, so that the model meets the balance requirements among the identification accuracy, response time and missed identification rate, including:

[0130] A normalized benchmark value of the response time and the missed identification rate is set;

[0131] The actual data of the response time and the missed identification rate generated by the model identification are normalized based on the normalized benchmark value;

[0132] The identification accuracy of the model is weighted with the normalized response time and missed identification rate according to a preset weight coefficient to obtain a comprehensive performance result;

[0133] The model is continuously adjusted and trained with the optimal comprehensive performance result as the target until the comprehensive performance meets the standard or the iteration number reaches the upper limit, so as to complete the dynamic optimization.

[0134] In addition, Figure 2 The detailed work processes of the communication interaction unit, the emergency angle coordination unit and the algorithm self-upgrading unit can refer to the design in the above method embodiments, and will not be repeated here.

[0135] Those skilled in the art can clearly understand that, for the convenience and brevity of description, only the division of the above functional units and modules is exemplified, and in actual application, the above functions can be completed by different functional units and modules according to needs, that is, the internal structure of the system is divided into different functional units or modules to complete all or part of the functions described above. Each functional unit or module in the embodiment can be integrated in one processing unit, or each unit can be physically present separately, or two or more units can be integrated in one unit. The integrated unit can be realized in the form of hardware or in the form of software functional unit. In addition, the specific names of each functional unit or module are only for the convenience of mutual distinction, and do not limit the protection scope of the present application. The specific working process of the unit or module in the system can refer to the corresponding process in the foregoing method embodiment, which will not be repeated here.

[0136] With reference to Figure 3 The embodiment of the present application also provides a computer device, comprising a memory and a processor and a computer program stored in the memory, when the computer program is executed on the processor, the deep learning based power grid hidden danger identification method is realized.

[0137] The computer device can be a desktop computer, a notebook computer, a palm computer and a cloud server and the like. The computer device can include, but is not limited to, a processor and a memory. Those skilled in the art can understand that, Figure 3 It is only an example of the computer device, and does not constitute a limitation on the computer device, and can include more or fewer components than the illustration, or combine certain components, or different components, for example, it can also include an input and output device, a network access device and the like.

[0138] The processor can be a central processing unit (CPU), and the processor can also be other general-purpose processors, digital signal processors (DSP), application specific integrated circuits (ASIC), field programmable gate arrays (FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor.

[0139] The memory can be an internal storage unit of the computer device in some embodiments, such as a hard disk or a memory of the computer device. The memory can also be an external storage device of the computer device in other embodiments, such as a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card, and the like. Further, the memory can include both an internal storage unit and an external storage device of the computer device. The memory is used to store an operating system, an application program, a boot loader, data, and other programs, such as program codes of the computer program. The memory can also be used to temporarily store data that has been output or will be output.

[0140] The embodiment of the present application further provides a computer readable storage medium, which stores a computer program. The computer program is run by a processor to implement the power grid hidden danger identification method based on deep learning according to any one of the above methods.

[0141] In the embodiment, the integrated unit, if realized in the form of a software function unit and sold or used as an independent product, can be stored in a computer readable storage medium. Based on such understanding, the present application can implement all or part of the processes of the above-mentioned embodiment methods by a computer program to instruct related hardware, and the computer program can be stored in a computer readable storage medium. The computer program can implement the steps of each method embodiment when executed by a processor. The computer program includes computer program codes, which can be in the form of source code, object code, executable files, or some intermediate forms, etc. The computer readable medium at least includes any entity or device capable of carrying the computer program code to a photographing device / terminal device, a recording medium, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium. For example, a U disk, a mobile hard disk, a magnetic disk or an optical disk, etc. In some jurisdictions, according to legislation and patent practice, the computer readable medium can not be an electrical carrier signal and a telecommunication signal.

[0142] The embodiment of the present application provides a computer program product, which includes a computer program. The computer program is executed by a processor to implement the power grid hidden danger identification method based on deep learning according to any one of the above methods.

[0143] In the above embodiments, the description of each embodiment is focused on, and the part not described or recorded in a certain embodiment can be referred to the relevant description of other embodiments.

[0144] Those skilled in the art can understand that the units and algorithm steps of each example described in combination with the embodiments disclosed herein can be realized by electronic hardware or a combination of computer software and electronic hardware. Whether the functions are realized in hardware or software depends on the specific application and design constraints of the technical solution. A person skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0145] In the embodiments disclosed in the present application, it should be understood that the disclosed apparatus / terminal device and method can be implemented in other ways. For example, the apparatus / terminal device embodiments described above are only schematic, and the division of the modules or units is only a logical function division, and there can be another division in actual implementation. For example, a plurality of units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the displayed or discussed mutual coupling or direct coupling or communication connection between each other can be indirect coupling or communication connection through some interface, device or unit, and can be electrical, mechanical or in other forms.

[0146] The above embodiments are only used to illustrate the technical solutions of the present application, but not limit it; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that the technical solutions recorded in the foregoing embodiments can be modified, or some technical features can be replaced by equivalent; and these modifications or replacements do not make the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A deep learning-based method for identifying potential power grid hazards, characterized in that, The power grid hazard identification model is based on a pre-trained model. The model is used to identify different types of power grid hazards. During the pre-training process, the model is dynamically optimized based on the identification results to ensure that the model meets the balance requirements between identification accuracy, response time and missed identification rate. The method includes: Acquire multi-dimensional imagery data of the power grid from air, space, and ground; the multi-dimensional imagery data from air, space, and ground is cross-spatial hierarchical data covering airspace, sky, and land. The air-space-ground multi-dimensional image data is standardized and preprocessed to obtain standardized data. Local detail features within the region and global correlation features across the region are extracted from the standardized data. The extracted features are then fused to obtain multi-dimensional fused feature data. The multi-dimensional fused feature data is input into the trained power grid hazard identification model to obtain the hazard identification result.

2. The deep learning-based power grid hazard identification method according to claim 1, characterized in that, The aforementioned multi-dimensional air-space imagery data includes at least: Satellite imagery in the sky, drone imagery in the airspace, and ground-based infrared sensor and surveillance imagery in the ground.

3. The deep learning-based power grid hazard identification method according to claim 1, characterized in that, The aforementioned air-space-ground multi-dimensional image data undergoes standardized preprocessing to obtain standardized data, including: The format and pixel arrangement rules of the aforementioned air-space-ground multi-dimensional image data are unified to obtain the first image data; The second image data is obtained by performing scale matching on image data of different resolutions in the first image data using a bilinear interpolation algorithm. The second image data is subjected to radiometric correction, geometric correction, and fusion processing to obtain the standardized data.

4. The deep learning-based power grid hazard identification method according to claim 1, characterized in that, The standardized data is subjected to local detail feature extraction within a region and global correlation feature extraction across regions, including: Local detail features within the region are extracted using a convolutional neural network. The extracted local detail features include at least the tower tilt angle, the outline of the hanging objects on the conductor, the area of ​​water accumulation on the ground, and the area of ​​abnormal heat generation of the equipment. The global correlation features across regions are extracted using the Transformer architecture. The extracted global correlation features include at least the spatial distance between buildings and power grid lines, cross-regional terrain change correlations, and disaster scenario spread trends.

5. The deep learning-based power grid hazard identification method according to claim 1, characterized in that, During the pre-training process, the power grid hazard identification model is dynamically optimized based on the identification results to ensure that the model meets the balance requirements among identification accuracy, response time, and missed identification rate, including: Set the normalized baseline value for the response time and the missed detection rate; The actual data of response time and missed identification rate generated by the model identification are normalized based on the normalized benchmark value; The model's recognition accuracy is weighted by the normalized response time and the missed recognition rate according to the preset weight coefficients to obtain the comprehensive performance result. With the goal of achieving the best overall performance, the model is continuously adjusted and trained until the overall performance meets the standard or the number of iterations reaches the upper limit, thus completing dynamic optimization.

6. A deep learning-based power grid hazard identification system, characterized in that, The power grid hazard identification model is based on a pre-trained model. The model is used to identify different types of power grid hazards. During the pre-training process, the model is dynamically optimized based on the identification results to ensure that the model meets the balance requirements between identification accuracy, response time and missed identification rate. The system includes: A multi-source image acquisition unit is used to acquire multi-dimensional image data of the power grid in the air, space, and ground; the multi-dimensional image data in the air, space, and ground is cross-spatial hierarchical data covering the airspace, sky, and land. The deep learning processing unit is used to perform standardized preprocessing on the air-space-ground multi-dimensional image data to obtain standardized data, and to extract local detail features within the region and global correlation features across the region from the standardized data, and to fuse the extracted features to obtain multi-dimensional fused feature data. The power grid hazard identification unit is used to input the multi-dimensional fused feature data into the trained power grid hazard identification model to obtain the hazard identification result.

7. The power grid hazard identification system based on deep learning according to claim 6, characterized in that, In the multi-source image acquisition unit, the air-space-ground multi-dimensional image data includes at least: Satellite imagery in the sky, drone imagery in the airspace, and ground-based infrared sensor and surveillance imagery in the ground.

8. The power grid hazard identification system based on deep learning according to claim 6, characterized in that, The deep learning processing unit performs standardized preprocessing on the air-space-ground multi-dimensional image data to obtain standardized data, including: The format and pixel arrangement rules of the aforementioned air-space-ground multi-dimensional image data are unified to obtain the first image data; The second image data is obtained by performing scale matching on image data of different resolutions in the first image data using a bilinear interpolation algorithm. The second image data is subjected to radiometric correction, geometric correction, and fusion processing to obtain the standardized data.

9. The power grid hazard identification system based on deep learning according to claim 6, characterized in that, In the deep learning processing unit, local detail feature extraction within a region and global correlation feature extraction across regions are performed on the standardized data, including: Local detail features within the region are extracted using a convolutional neural network. The extracted local detail features include at least the tower tilt angle, the outline of the hanging objects on the conductor, the area of ​​water accumulation on the ground, and the area of ​​abnormal heat generation of the equipment. The global correlation features across regions are extracted using the Transformer architecture. The extracted global correlation features include at least the spatial distance between buildings and power grid lines, cross-regional terrain change correlations, and disaster scenario spread trends.

10. The power grid hazard identification system based on deep learning according to claim 6, characterized in that, During the pre-training process, the power grid hazard identification model is dynamically optimized based on the identification results to ensure that the model meets the balance requirements among identification accuracy, response time, and missed identification rate, including: Set the normalized baseline value for the response time and the missed detection rate; The actual data of response time and missed identification rate generated by the model identification are normalized based on the normalized benchmark value; The model's recognition accuracy is weighted by the normalized response time and the missed recognition rate according to the preset weight coefficients to obtain the comprehensive performance result. With the goal of achieving the best overall performance, the model is continuously adjusted and trained until the overall performance meets the standard or the number of iterations reaches the upper limit, thus completing dynamic optimization.

Images