Electric power operation unmanned aerial vehicle safety supervision target identification and detection method, system and device, and medium

Abstract

The invention discloses an electric power operation unmanned aerial vehicle safety supervision target identification and detection method, system and device, and a medium. The method comprises the following steps: carrying out electric power operation risk identification based on environment characteristics and task attributes of an electric power operation area; configuring a monitoring frequency, and performing data acquisition on a power operation area; constructing a target region segmentation device based on a predetermined operation monitoring target, performing target segmentation on the monitoring image, and outputting a region image set; traversing and comparing the regional image set acquired under the current monitoring node with the historical regional image set acquired under the previous adjacent monitoring node, selecting images with significant differences, and performing point cloud fitting to construct a three-dimensional target simulation model; and building an anomaly identification plug-in based on a predetermined monitoring index and the operation early warning area, carrying out anomaly detection on the three-dimensional target simulation model, and outputting an operation detection result. According to the scheme, global intelligent monitoring of electric power high-altitude operation is achieved, potential safety hazards are found in time, and the timeliness and accuracy of risk identification are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of power operation safety monitoring, and particularly to a method, system, device and medium for identifying and detecting safety supervision targets of unmanned aerial vehicles for power operations. Background Art

[0002] As an important part of the daily operation and maintenance of the power system, high-altitude power operations involve multiple high-altitude operation sites such as transmission lines, substations, and wind turbines. Due to the particularity and danger of the operation environment, the safety management of high-altitude power operations has always been the focus of attention of power companies and operation units.

[0003] Currently, traditional high-altitude power operation safety monitoring methods mainly rely on manual inspections, ground monitoring equipment, and fixed cameras, etc. It is difficult to fully cover all areas of the operation site, and it is impossible to monitor the operation site in real time and all-round. There are technical problems such as slow response speed, insufficient monitoring comprehensiveness and accuracy. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide a method, system, device and medium for identifying and detecting safety supervision targets of unmanned aerial vehicles for power operations in view of the technical problems of the traditional high-altitude power operation safety monitoring method, which cannot achieve real-time monitoring in real time and all-round, and has slow response speed, insufficient monitoring comprehensiveness and accuracy.

[0005] To achieve the above object, the technical solution provided by the present invention is as follows:

[0006] The first aspect of the present application provides a method for identifying and detecting safety supervision targets of unmanned aerial vehicles for power operations, including the following steps:

[0007] S1: Identify power operation risks based on the environmental characteristics and task attributes of the power operation area, and determine the operation risk coefficient and operation warning area;

[0008] S2: Configure the monitoring frequency according to the operation risk coefficient, and based on a predetermined monitoring plan, control the unmanned aerial vehicle to collect data from the power operation area at the monitoring frequency to obtain monitoring images and a monitoring point cloud data set, wherein the monitoring frequency is positively correlated with the operation risk coefficient;

[0009] S3: Set the monitoring time node according to the actual scenario, and construct a target area segmenter based on a predetermined operation monitoring target. Use the target area segmenter to segment the monitoring image to output a set of regional images;

[0010] S4: Traverse and compare the set of regional images collected under the current monitoring node with the set of historical regional images collected under the previous adjacent monitoring node. Select the regional images with a similarity less than the preset similarity threshold as significantly different images. Based on the obtained set of significantly different images, perform point cloud fitting in the monitoring point cloud dataset respectively, and construct a three-dimensional target simulation model according to the point cloud fitting results;

[0011] S5: Build an anomaly recognition plugin based on the predetermined monitoring indicators and the operation warning area, and use the anomaly recognition plugin to detect anomalies in the three-dimensional target simulation model, and output the operation detection results.

[0012] To optimize the above technical solution, the specific measures taken also include:

[0013] Step S1 includes:

[0014] Taking the regional characteristics of the power operation area as a constraint, retrieve and obtain the sample environmental feature set and the sample task attribute set, and have domain experts mark the operation risk coefficient and the operation warning area for them to obtain the sample risk coefficient set and the sample warning area set;

[0015] Using the sample environmental features and sample task attributes as inputs, and the sample risk coefficient set and the sample warning area set as supervision, perform supervised training with a BP neural network until convergence to obtain an operation risk recognition channel for identifying and determining the operation risk coefficient and the operation warning area.

[0016] Step S3 includes:

[0017] Obtain the predetermined operation monitoring target. Taking the regional characteristics, environmental characteristics, and task attributes as retrieval constraints, collect sample operation images, and perform annotation and bounding on different sample operation images according to the predetermined operation monitoring target to obtain the sample regional image set;

[0018] Using the sample operation images and the sample regional image set as training data, perform supervised training on the convolutional neural network until convergence to generate a target region segmenter for target segmentation of the monitoring images.

[0019] Step S4 includes:

[0020] Taking the regional characteristics, environmental characteristics, task attributes, and the predetermined operation monitoring target as retrieval constraints, collect the sample regional image set. Each group of regional images in the sample regional image set is two images containing the same monitoring target, and perform similarity annotation on the two images containing the same monitoring target;

[0021] Use the sample regional image set and the marked similarity to train a siamese network to obtain a power image comparison model that meets convergence;

[0022] Using the power image comparison model, traverse and compare the regional image set with the historical regional image set of the previous adjacent monitoring node, and output the image similarity.

[0023] According to a preset similarity threshold, screen the obtained image similarity, select the regional images with similarity less than the preset similarity threshold as significantly different images, and output the significantly different image set.

[0024] Further, step S4 further includes:

[0025] Randomly select a first significantly different image from the significantly different image set, and extract the contour features of the first significantly different image, and output the first contour features.

[0026] Based on the first contour features, perform point cloud random fitting in the monitoring point cloud dataset to obtain a first point cloud fitting result, and calculate a first fitting degree; wherein, the first fitting degree is the ratio of the number of point clouds falling within the first contour features to the total number of point clouds in the monitoring point cloud dataset.

[0027] Based on the first contour features, perform the next point cloud fitting in the monitoring point cloud dataset to obtain a second point cloud fitting result and a second fitting degree.

[0028] Continue iterative fitting until a predetermined number of times is reached, output multiple point cloud fitting results and multiple fitting degrees, and select the point cloud fitting result with the largest fitting degree as the first optimal point cloud fitting result.

[0029] Successively obtain the optimal point cloud fitting results of all significantly different images in the significantly different image set, and respectively simulate and construct three-dimensional target simulation models according to the respective optimal point cloud fitting results.

[0030] Further, step S5 includes:

[0031] Configure predetermined monitoring indicators corresponding to the predetermined operation monitoring target.

[0032] Randomly select a first monitoring target from the predetermined operation monitoring target, and obtain the first monitoring indicator of the first monitoring target.

[0033] Based on the first monitoring target, retrieve and obtain a first sample three-dimensional target model from the historical three-dimensional model database, and perform normal / abnormal annotation on the first sample three-dimensional target model according to the first monitoring indicator and the operation warning area, to obtain a first sample annotation result.

[0034] Use the first sample 3D target model and the first sample annotation results as training data, randomly divide the training data into Q parts, construct Q training sets through sampling with replacement, construct the first training set, and iterate Q times to obtain Q training sets;

[0035] Use the Q training sets to perform supervised training on the convolutional neural network to obtain Q first anomaly recognition branches, and construct the first anomaly recognition unit based on the Q first anomaly recognition branches;

[0036] Successively obtain the anomaly recognition units corresponding to each monitoring target in the predetermined job monitoring target, and fuse and construct an anomaly recognition plug-in based on the mapping and association relationship between the monitoring target and the anomaly recognition unit.

[0037] Further, the method for fusing and constructing an anomaly recognition plug-in based on the mapping and association relationship between the monitoring target and the anomaly recognition unit is as follows:

[0038] Calculate the similarity deviation between each image similarity and the preset similarity threshold respectively;

[0039] Calculate the ratio of the similarity deviation to the corresponding historical maximum similarity deviation respectively to determine each deviation scale coefficient, where each deviation scale coefficient corresponds to each monitoring target one by one;

[0040] Multiply each deviation scale coefficient by Q to obtain multiple branch call numbers;

[0041] Based on the branch call numbers, randomly select the same number of anomaly recognition branches in multiple anomaly recognition units of the anomaly recognition plug-in to generate an adapted anomaly recognition plug-in;

[0042] Use the adapted anomaly recognition plug-in to perform anomaly detection on the 3D target simulation model and output the job detection results of the monitoring target.

[0043] The second aspect of this application provides a power operation UAV safety supervision and inspection target recognition and detection system, including:

[0044] A job risk recognition module for performing power operation risk recognition based on the environmental characteristics and task attributes of the power operation area to determine the job risk coefficient and the job warning area;

[0045] A job data acquisition module for configuring the monitoring frequency according to the job risk coefficient, and controlling the UAV to perform data acquisition according to the monitoring frequency based on a predetermined monitoring plan to obtain monitoring images and a monitoring point cloud data set;

[0046] A monitoring image segmentation module for constructing a target area segmenter based on a predetermined job monitoring target, and using the target area segmenter to perform target segmentation on the monitoring image to output a regional image set;

[0047] A simulation model construction module is configured to traverse and compare the regional image set with the historical regional image set of the previous adjacent monitoring node, select the regional images with similarity less than a preset similarity threshold as significantly different images, and respectively perform point cloud fitting in the monitoring point cloud data set based on the significantly different image set, and construct a three-dimensional target simulation model according to the point cloud fitting result;

[0048] A target anomaly detection module is configured to build an anomaly recognition plug-in based on a predetermined monitoring index and the operation warning area, and use the anomaly recognition plug-in to perform anomaly detection on the three-dimensional target simulation model, and output an operation detection result.

[0049] The third aspect of the present application provides an electronic device, including: a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the computer program, the above-mentioned power operation UAV safety supervision and inspection target recognition and detection method is implemented.

[0050] The fourth aspect of the present application provides a computer-readable storage medium storing a computer program, and the computer program causes a computer to execute the above-mentioned power operation UAV safety supervision and inspection target recognition and detection method.

[0051] Compared with the prior art, the beneficial effects of the present invention are:

[0052] The present invention first implements risk identification based on the environmental characteristics and task attributes of the power operation area to generate an operation risk coefficient and a warning area; secondly, dynamically configures the UAV monitoring frequency according to the risk coefficient, and obtains monitoring images and point cloud data sets based on a preset scheme, where the monitoring frequency is positively correlated with the risk level; then constructs a target area segmenter to process the monitoring images and outputs a refined regional image set; then extracts a set of different images with similarity lower than the threshold by comparing the current and historical regional images in full quantity, and constructs a three-dimensional simulation model based on the corresponding point cloud data; finally, combines the monitoring index and the warning area to construct an anomaly recognition algorithm and outputs multi-dimensional detection results.

[0053] This solution realizes the full-domain intelligent monitoring of high-altitude power operations, timely discovers potential safety hazards, significantly improves the timeliness and accuracy of risk identification, can shorten the warning time for identifying hazards, synchronously optimize the operation efficiency, and effectively ensure personnel safety and construction quality. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 It is a schematic flowchart of a UAV safety supervision and inspection target recognition and detection method provided by the present invention.

[0055] Figure 2Schematic diagram of a drone safety supervision and inspection target recognition and detection system provided by the present invention.

[0056] Figure 3 Schematic diagram of an electronic device provided by the present invention.

[0057] Figure 4 Schematic diagram of a computer-readable storage medium provided by the present invention.

[0058] In the drawings, the components represented by the reference numerals are described as follows:

[0059] Operation risk identification module 01, operation data acquisition module 02, monitoring image segmentation module 03, simulation model construction module 04, target anomaly detection module 05, electronic device 500, memory 510, processor 520, first computer program 511, computer-readable storage medium 600, second computer program 611. Detailed implementation manners

[0060] The above content of the present invention will be further described in detail below in the form of specific implementation manners. However, it should not be understood that the scope of the above subject matter of the present invention is limited to the following embodiments. Any technology implemented based on the above content of the present invention belongs to the scope of the present invention.

[0061] Embodiment 1

[0062] As Figure 1 shown, the embodiment of the present invention provides a method for drone safety supervision and inspection target recognition and detection, which specifically includes the following steps:

[0063] S100: Identify power operation risks based on the environmental characteristics and task attributes of the power operation area, and determine the operation risk coefficient and operation warning area.

[0064] Furthermore, step S100 of the present invention further includes:

[0065] S110: Read the environmental characteristics and task attributes of the power operation area. Among them, the environmental characteristics at least include wind force, wind direction, brightness, and temperature, and the task attributes at least include task type and operation height; S120: With the regional characteristics of the power operation area as a constraint, retrieve and obtain a sample environmental characteristic set and a sample task attribute set, and use domain experts to identify the operation risk coefficients and operation warning areas under different sample environmental characteristics and different sample task attributes, and obtain a sample risk coefficient set and a sample warning area set; S130: Use the sample environmental characteristics and sample task attributes as inputs, and the sample risk coefficients and sample warning areas as supervision, and use the sample environmental characteristic set, sample task attribute set, sample risk coefficient set, and sample warning area set to supervise and train the BP neural network until convergence to obtain an operation risk identification channel; S140: Input the environmental characteristics and task attributes into the operation risk identification channel, and output the operation risk coefficient and the operation warning area.

[0066] Specifically, first, read the environmental characteristics and task attributes of the power operation area. Among them, the environmental characteristics refer to the physical environmental parameters of the operation site, which have important impacts on the safety, efficiency of the operation, and the health and safety of the operation personnel. Common environmental characteristics include wind force, wind direction, brightness (the illumination intensity of the operation area), and temperature. Temperature will directly affect the working comfort of the operation personnel, the operating conditions of the equipment, and safety protection. Too high or too low temperature may increase the accident risk. The task attributes include task type, operation height, etc. The task type refers to the specific nature or category of the operation, such as installation, maintenance, cleaning, inspection, etc. Different types of tasks correspond to different operation requirements, tools, equipment, and safety measures; the operation height refers to the vertical height from the working platform or operation point to the ground when the operation personnel perform the task. The operation height directly affects the safety risk of the operation personnel. The higher the operation area, the higher the risk usually is.

[0067] Next, obtain the regional characteristics of the power operation area. The regional characteristics include the operation area type, operation area height, and operation area structure (such as the stability of the power tower, the layout of the overhead line, etc.). Among them, different types of operation areas have different safety requirements and risks. For example, power towers, overhead line poles, wind turbine towers, substations, etc. all belong to different types of operation areas. Then, use the regional characteristics of the power operation area as a constraint for information retrieval, obtain the matching environmental characteristic set and task attribute set, and get the sample environmental characteristic set and sample task attribute set. Further, combine the experience and knowledge of domain experts to label different sample data (different sample environmental characteristics and different sample task attributes), identify the operation risk coefficients and operation warning areas in each operation scenario, and obtain the sample risk coefficient set and sample warning area set.

[0068] Then, a job risk identification channel is constructed based on a BP neural network. The job risk identification channel is used to analyze the job risk coefficient and job warning area under different job scenarios, including an input layer, multiple hidden layers, and an output layer. The input data of the input layer are environmental characteristics and task attributes, and the output data of the output layer are the risk coefficient and warning area. Next, with the sample environmental characteristics and sample task attributes as inputs and the sample risk coefficient and sample warning area as supervision, the job risk identification channel is supervised and trained using the sample environmental characteristics set, sample task attributes set, sample risk coefficient set, and sample warning area set. First, the input data (environmental characteristics and task attributes) are input into the input layer of the neural network. The input data are weighted and summed through each hidden layer and undergo a non-linear transformation through an activation function (such as Sigmoid or ReLU), and finally are transmitted to the output layer. The output layer generates the job risk coefficient and job warning area. Then, the error between the network output result and the target data (sample risk coefficient and warning area) is calculated through a loss function (such as mean squared error), and the error is transmitted back to each layer through backpropagation. The gradient of the error with respect to the weights and biases is calculated layer by layer, and the gradient descent algorithm (such as Stochastic Gradient Descent SGD) is used to update the weights and biases of each neuron. The above steps are repeated for iterative training. By continuously adjusting the weights and biases, the error between the network output and the target output is gradually reduced until the loss function converges or reaches a preset stopping condition (such as the maximum number of iterations), and the trained job risk identification channel is outputted.

[0069] Finally, the environmental characteristics and task attributes are input into the job risk identification channel for risk identification, and the job risk coefficient and job warning area are output. Among them, the larger the job risk coefficient, the higher the degree of job risk.

[0070] S200: Configure the monitoring frequency according to the job risk coefficient. Based on a predetermined monitoring plan, control the drone to collect data on the power operation area according to the monitoring frequency, and obtain the monitoring images and monitoring point cloud data sets. Among them, the monitoring frequency is positively correlated with the job risk coefficient.

[0071] Specifically, the monitoring frequency is configured according to the operation risk coefficient. Among them, the monitoring frequency is positively correlated with the operation risk coefficient, that is, the greater the operation risk coefficient, the higher the monitoring frequency. By setting the dynamic monitoring frequency according to the operation risk level, it can ensure that monitoring data can be obtained more frequently and timely in a high-risk operation environment, ensuring operation safety. For example, for high-risk operation scenarios (such as high wind speed, high temperature, high altitude operations, etc.), the operation risk coefficient is relatively large. In order to more accurately monitor the operation environment and timely discover potential safety hazards, the monitoring frequency will be increased accordingly. Data collection can be carried out by drones at shorter time intervals (such as obtaining image and point cloud data every second or every few seconds); for low-risk operation scenarios (such as moderate wind, stable environment, etc.), the operation risk coefficient is relatively small, and the monitoring frequency can be appropriately reduced, and data collection can be carried out every longer time (such as every minute or every few minutes).

[0072] Configure a predetermined monitoring plan based on the actual monitoring scenario, environmental characteristics, and task attributes, including the drone flight altitude, monitoring equipment, data collection accuracy, etc.; according to the predetermined monitoring plan and monitoring frequency, control the drone to collect data in the power operation area to obtain monitoring images and a monitoring point cloud data set.

[0073] By configuring the monitoring frequency according to the operation risk coefficient, the drone can perform reasonable data collection at different risk levels, thereby achieving comprehensive and accurate real-time monitoring of the operation environment. In high-risk scenarios, frequent monitoring helps to timely discover potential problems and take measures, while in low-risk scenarios, appropriately reducing the monitoring frequency can save resources while ensuring safety.

[0074] S300: Set the monitoring time node according to the actual scenario, and construct a target area segmenter based on the predetermined operation monitoring target. Use the target area segmenter to perform target segmentation on the monitoring image and output a set of regional images.

[0075] Furthermore, step S300 of the present invention further includes:[[]]

[0076] S310: Obtain the predetermined operation monitoring target, where the predetermined operation monitoring target at least includes the operation location, operation components, operation tools, and operation personnel; S320: Using regional features, environmental features, and task attributes as retrieval constraints, collect multiple sample operation images, and perform annotation bounding on different sample operation images according to the predetermined operation monitoring target to obtain multiple sample regional image sets; S330: Using multiple sample operation images and multiple sample regional image sets as training data, perform supervised training on the convolutional neural network until convergence to generate a target area segmenter; S340: Input the monitoring image into the target area segmenter and output a set of regional images.

[0077] Specifically, first, obtain a predetermined job monitoring target. The predetermined job monitoring target at least includes a job location, job components, job tools, and job personnel. The job location refers to the area where actual operations are carried out in high-altitude operations, usually the specific location where the equipment is located (such as a transmission tower, a certain power equipment in a substation, the blade of a wind turbine, etc.); job components refer to all equipment and components involved in high-altitude power operations (such as wires on a power tower, electrical switches in a substation, the blade of a wind turbine, etc.); job tools include all equipment and tools used in high-altitude power operations (such as wrenches, insulating rods, safety belts, hooks, etc.); job personnel refer to the personnel participating in high-altitude power operations, especially whether they operate in accordance with safety regulations (such as whether they wear safety belts, whether they comply with the operation procedures, etc.).

[0078] Then, perform information retrieval with regional characteristics, environmental characteristics, and task attributes as constraints. For example, retrieve historical job logs and collect multiple sample job images; then, perform annotation and bounding box selection on different sample job images according to the predetermined job monitoring target. For example, domain experts or operators manually mark the areas of the job area, job components, tools, and job personnel based on the collected images, and perform bounding box selection and cropping on the marked areas. Usually, a rectangular box is used to represent the spatial position of the target in the bounding box selection, and each rectangular box will be marked with the corresponding category (job location, tool, personnel, etc.) and the precise coordinates of the area, obtaining a set of multiple sample regional images, and each sample regional image contains different monitoring targets (job location, components, tools, personnel, etc.).

[0079] Then, construct a target region segmenter based on a convolutional neural network. The target region segmenter is used to automatically identify and segment the target regions in high-altitude power operation images, including an input layer, a convolutional layer, a pooling layer, a fully connected layer, and an output layer. Among them, the input data of the input layer is the job image, and the output data of the output layer is the set of segmented regional images; then, use multiple sample job images and multiple sample regional image sets as training data to perform supervised training on the convolutional neural network. First, the input image undergoes forward propagation through the network, and the convolutional layer extracts features, and finally outputs a segmentation result to identify the target regions in the job image; then, by comparing the difference between the prediction result and the true label (i.e., the target region selected by annotation and bounding box), use the cross-entropy loss function to calculate the loss value; then, according to the calculated loss value, calculate the gradient of each layer through the backpropagation algorithm and update the weight parameters in the network. This process optimizes the network through gradient descent (such as the Adam optimizer) to enable the network to learn the optimal weights and biases; repeat the above process, gradually adjust the network parameters until the model converges (such as the loss function tends to be minimized) to achieve the expected target region segmentation effect, and obtain the trained target region segmenter.

[0080] Finally, the monitored images are input into the target area segmenter for area division, and a set of area images is output. By constructing the target area segmenter based on a convolutional neural network for image segmentation, the intelligence, efficiency, and accuracy of image segmentation can be improved, providing strong support for the efficient monitoring of high-altitude power operations.

[0081] S400: Traverse and compare the set of area images collected under the current monitoring node with the historical set of area images collected under the previous adjacent monitoring node. Select the area images with a similarity less than the preset similarity threshold as significantly different images. Based on the obtained set of significantly different images, perform point cloud fitting in the monitoring point cloud dataset respectively, and construct a three-dimensional target simulation model according to the point cloud fitting results.

[0082] Furthermore, step S400 of the present invention further includes:

[0083] S410: Using area features, environmental features, task attributes, and a predetermined operation monitoring target as retrieval constraints, collect multiple groups of sample area images, and label to obtain multiple sample image similarities. Each group of area images is two images containing the same monitoring target; S420: Train a siamese network using multiple groups of sample area images and multiple sample image similarities to obtain a converged power image comparison model; S430: Use the power image comparison model to traverse and compare the set of area images with the historical set of area images of the previous adjacent monitoring node, and output multiple image similarities; S440: Screen the multiple image similarities according to the preset similarity threshold, and select the area images with a similarity less than the preset similarity threshold as significantly different images, and output the set of significantly different images.

[0084] Specifically, first, perform information retrieval with area features, environmental features, task attributes, and a predetermined operation monitoring target as constraints, and collect multiple groups of sample area images. Each group of sample area images is two images containing the same monitoring target (such as the same operation component, the same operator, etc.); then calculate and label the similarity of each group of sample area images respectively. For example, through the Euclidean distance algorithm, extract image features and calculate the Euclidean distance between feature vectors. The smaller the distance, the more similar the images are considered, and multiple sample image similarities are obtained.

[0085] Next, a power image comparison model is built based on the Siamese network. The Siamese network is a deep learning model composed of two identical sub-networks (with shared weights), which is widely used to solve tasks such as image comparison, image similarity calculation, and contrast learning. This model is trained by learning the similarity or difference between two input images. The power image comparison model includes an input layer, a convolutional layer, a similarity comparison layer, and an output layer. The convolutional layer can identify low-level features such as edges, textures, and shapes in the image. Then, multiple groups of sample region images and multiple sample image similarities are used as training data to perform supervised training on the power image comparison model. First, in the training stage, the sample region images and similarity labels are input into the Siamese network. The network generates predicted similarity scores through forward propagation and compares them with the actual labels (0 or 1). Then, the loss is calculated and the weights of the network are optimized through backpropagation. Through multiple iterative trainings, the Siamese network will learn to extract and compare image features and finally converge.

[0086] Then, the historical region image set of the previous adjacent monitoring node is obtained, and the region image set and the historical region image set are integrated to obtain multiple input image groups. Each input image group consists of two images, one from the region image at the current moment and the other from the region image at the historical moment. By pairing these two images, the Siamese network can perform similarity analysis to evaluate whether there are significant changes (such as safety measures, equipment status, etc.) between the two images. Then, the power image comparison model is used to perform similarity recognition on multiple input image groups respectively, and multiple image similarities are output. By using the power image comparison model to perform similarity recognition on multiple input image groups, the difference between the current operation site and the historical operation state can be accurately evaluated, which can not only timely discover potential safety hazards, but also provide intelligent, safe and reliable real-time monitoring support for high-altitude power operations, improving the safety and efficiency of operations.

[0087] According to factors such as the operation risk coefficient, operation environment, and task requirements, a similarity threshold (usually a value between 0 and 1) is preset. Images with a similarity higher than this threshold are considered to have no significant changes and belong to the normal operation state. Images with a similarity lower than this threshold indicate that there have been significant changes or abnormalities at the operation site, and there may be potential safety hazards. Then, multiple image similarities are screened according to the preset similarity threshold, and the region images with similarities less than the preset similarity threshold are selected as significantly different images. Significantly different images indicate that there are obvious differences between the current operation state and the historical state, which may be caused by equipment failures, safety hazards, or other abnormal situations, and a set of significantly different images is obtained.

[0088] Furthermore, step S400 of the present invention further includes:

[0089] S450: Randomly select the first significantly different image from the set of significantly different images, extract the contour features of the first significantly different image, and output the first contour feature; S460: Based on the first contour feature, perform random point cloud fitting in the monitored point cloud dataset to obtain the first point cloud fitting result, and calculate the first fitting degree, where the first fitting degree is the ratio of the number of point clouds falling within the first contour feature to the total number of point clouds in the monitored point cloud dataset; S470: Again, according to the first contour feature, perform random point cloud fitting in the monitored point cloud dataset to obtain the second point cloud fitting result and the second fitting degree; S480: Continue iterative fitting until the predetermined number of times is reached, output multiple point cloud fitting results and multiple fitting degrees, and select the point cloud fitting result with the maximum fitting degree as the first optimal point cloud fitting result; S490: Sequentially obtain the multiple optimal point cloud fitting results of multiple significantly different images in the set of significantly different images, and simulate and construct multiple three-dimensional target simulation models based on the multiple optimal point cloud fitting results.

[0090] Specifically, first, randomly select any significantly different image from the set of significantly different images and set it as the first significantly different image; then extract the contour features of the first significantly different image, that is, through image processing algorithms (such as edge detection, contour extraction, etc.), extract the contour features from the selected significantly different image. The contour features represent the edge, shape, and structural information of the object in the image, which can help the subsequent point cloud fitting work. Common contour extraction methods include Canny edge detection, Hough transform, etc., to obtain the first contour feature.

[0091] Next, based on the first contour feature, perform random point cloud fitting in the monitored point cloud dataset, that is, randomly distribute the point cloud data within the first contour feature to obtain the first point cloud fitting result, and calculate the fitting degree of the first point cloud fitting result. The fitting degree refers to the ratio of the number of point clouds falling within the range of the extracted contour feature to the total number of point clouds during the fitting process. The higher the fitting degree, the better the coincidence between the point cloud fitting result and the contour feature. Obtain the first fitting degree, where the first fitting degree is the ratio of the number of point clouds falling within the first contour feature to the total number of point clouds in the monitored point cloud dataset. Then, again, according to the first contour feature, perform random point cloud fitting in the monitored point cloud dataset to obtain the second point cloud fitting result and the second fitting degree. Use the same method to continue iterative fitting. By continuously performing random point cloud fitting, optimize the fitting process until the predetermined number of iterations is reached, output multiple point cloud fitting results and multiple fitting degrees, and select the point cloud fitting result with the maximum fitting degree as the first optimal point cloud fitting result.

[0092] Then, using the method of obtaining the same fitting result of the first optimal point cloud, the fitting results of multiple optimal point clouds of multiple significantly different images in the significantly different image set are obtained in sequence; finally, 3D simulation modeling is performed based on the fitting results of multiple optimal point clouds, that is, each fitting result of the optimal point cloud is converted into a 3D point cloud model, and 3D reconstruction is performed to construct multiple 3D target simulation models.

[0093] Point cloud fitting based on image contours is an effective 3D modeling technology that can combine the edge information in the image with point cloud data to accurately construct a 3D model of the high-altitude power operation scene, thereby improving the accuracy, accuracy, and reliability of subsequent target anomaly recognition.

[0094] S500: Build an anomaly recognition plugin based on predetermined monitoring indicators and operation warning areas, and use the anomaly recognition plugin to detect anomalies in the 3D target simulation model and output the operation detection results.

[0095] Furthermore, step S500 of the present invention further includes:

[0096] S510: Configure predetermined monitoring indicators, where the predetermined monitoring indicators correspond to predetermined operation monitoring targets and include position monitoring indicators, component monitoring indicators, tool monitoring indicators, and personnel safety indicators; S520: Randomly select a first monitoring target from the predetermined operation monitoring targets and obtain the first monitoring indicator of the first monitoring target; S530: Retrieve and obtain multiple first sample 3D target models based on the first monitoring target, and perform normal / anomaly labeling on the multiple first sample 3D target models according to the first monitoring indicator and the operation warning area to obtain multiple first sample labeling results; S540: Use the multiple first sample 3D target models and multiple first sample labeling results as training data, divide the training data into Q equal parts, select Q times with replacement, construct a first training set, iterate and select Q times to obtain Q training sets, where Q is an integer greater than 5; S550: Use the Q training sets to perform supervised training on the convolutional neural network to obtain Q first anomaly recognition branches, and construct a first anomaly recognition unit based on the Q first anomaly recognition branches; S560: Obtain multiple anomaly recognition units of multiple monitoring targets in sequence, and based on the mapping association between the monitoring targets and the anomaly recognition units, fuse and construct an anomaly recognition plugin according to the multiple anomaly recognition units of multiple monitoring targets.

[0097] Specifically, first, configure predetermined monitoring indicators. The predetermined monitoring indicators correspond to predetermined operation monitoring objectives and can be set according to the actual operation scenario, including position monitoring indicators (such as position deviation range, etc.), component monitoring indicators (such as component position, etc.), tool monitoring indicators (such as tool type, etc.), and personnel safety indicators (such as whether safety protection equipment is worn, including safety belts, helmets, gloves, etc.). Then, randomly select any one of the predetermined operation monitoring objectives as the first monitoring objective (such as an operator), and obtain the first monitoring indicator of the first monitoring objective (such as the personnel safety indicator). Then, based on the first monitoring objective, perform information retrieval to obtain multiple first sample three-dimensional target models, and respectively perform normal / abnormal annotation on the multiple first sample three-dimensional target models according to the first monitoring indicator and the operation warning area. Among them, satisfying the first monitoring indicator and not being in the operation warning area is a normal operation behavior, that is, there is no safety hazard during the operation process; on the contrary, if the first monitoring indicator is not satisfied and / or is in the operation warning area, it is an abnormal operation behavior, that is, there are potential safety hazards or non-compliance with operating procedures during the operation process; obtain multiple first sample annotation results.

[0098] Next, use the multiple first sample three-dimensional target models and the multiple first sample annotation results as training data, and divide the training data into Q equal parts, where Q is an integer greater than 5, and the value of Q can be set according to requirements, such as 10; then select Q times with replacement from the Q data sets to construct the first training set, and use the same method to iterate and select Q times to obtain Q training sets. Further, use the first sample three-dimensional target model as the input and the first sample annotation result (normal / abnormal) as the output, and use the Q training sets to perform supervised training on the convolutional neural network until the loss function converges to obtain Q first anomaly recognition branches. Each recognition branch is a convolutional neural network after supervised training, which can predict the behavior category (normal / abnormal) of the target based on the input three-dimensional target model, and construct the first anomaly recognition unit according to the Q first anomaly recognition branches.

[0099] Use the same method to sequentially obtain multiple anomaly recognition units (including Q recognition branches) for multiple monitoring objectives; then, based on the mapping association between the monitoring objectives and the anomaly recognition units, where each monitoring objective has a corresponding anomaly recognition unit. Through this mapping association, each monitoring objective can be associated with its corresponding anomaly recognition unit, so as to provide an accurate anomaly detection module for each monitoring objective, and fuse and construct an anomaly recognition plug-in according to the multiple anomaly recognition units of the multiple monitoring objectives.

[0100] Furthermore, step S500 of the present invention further includes:

[0101] S570: Calculate the similarity deviations between multiple image similarities and a preset similarity threshold respectively to obtain multiple similarity deviations; S580: Calculate the ratios of multiple similarity deviations to multiple historical maximum similarity deviations respectively to determine multiple deviation scale coefficients, where the deviation scale coefficients correspond one-to-one with the monitoring targets; S590: Multiply multiple deviation scale coefficients by Q respectively to obtain multiple branch call numbers; S5100: Based on multiple branch call numbers, randomly select the same number of anomaly recognition branches in multiple anomaly recognition units of the anomaly recognition plug-in to generate an adapted anomaly recognition plug-in; S5110: Use the adapted anomaly recognition plug-in to perform anomaly detection on multiple three-dimensional target simulation models and output multiple job detection results of multiple monitoring targets.

[0102] Specifically, first, calculate the similarity deviations between multiple image similarities and a preset similarity threshold respectively. The similarity deviation is the difference between the preset similarity threshold and the image similarity, and multiple similarity deviations are obtained. Then calculate the ratios of multiple similarity deviations to multiple historical maximum similarity deviations respectively, that is, calculate the ratio of the similarity deviation of the same monitoring target to the historical maximum similarity deviation (the maximum similarity deviation value collected during past monitoring can be used as a comparison standard), and multiple deviation scale coefficients are obtained. The deviation scale coefficients correspond one-to-one with the monitoring targets. Further, multiply multiple deviation scale coefficients by Q respectively and take the integer to obtain multiple branch call numbers, that is, the branch call number is proportional to the deviation scale coefficient. The larger the deviation scale coefficient, the more branch call numbers. By dynamically adjusting the number of branch calls in the anomaly recognition plug-in according to the similarity deviation size of the monitoring target, that is, for a monitoring target with a larger deviation (a situation where an anomaly may exist), the number of branch calls can be increased for more detailed and in-depth detection; for a monitoring target with a smaller deviation, the number of branch calls is reduced to avoid unnecessary calculations. By reducing the overprocessing of normal situations, the computing power resources of the system can be saved. Under normal circumstances, the deviation of the monitoring target is small, and not all branches need to be called, avoiding waste of computing resources and improving the analysis efficiency at the same time.

[0103] Then, according to multiple branch call numbers, randomly select the same number of anomaly recognition branches in multiple anomaly recognition units of the anomaly recognition plug-in to construct an adapted anomaly recognition plug-in. Finally, use the adapted anomaly recognition plug-in to perform anomaly detection on multiple three-dimensional target simulation models. The detection result is the mode output by multiple anomaly recognition branches (that is, the result with a large voting proportion), and multiple job detection results (normal or abnormal) of multiple monitoring targets are output.

[0104] The unmanned aerial vehicle safety supervision and inspection target recognition and detection method provided by the embodiment of the present invention has at least the following technical effects:

[0105] By identifying the risks of electric power operations based on the environmental characteristics and task attributes of the electric power operation area, determining the operation risk coefficient and the operation warning area; then configuring the monitoring frequency according to the operation risk coefficient, and controlling the drone to collect data according to the monitoring frequency based on a predetermined monitoring plan to obtain monitoring images and a monitoring point cloud data set, where the monitoring frequency is positively correlated with the operation risk coefficient; further constructing a target area segmenter based on a predetermined operation monitoring target, using the target area segmenter to perform target segmentation on the monitoring images, and outputting a set of regional images; then traversing and comparing the set of regional images with the historical set of regional images of the previous adjacent monitoring node, selecting the regional images with a similarity less than a preset similarity threshold as significantly different images, and respectively performing point cloud fitting in the monitoring point cloud data set based on the set of significantly different images, and constructing a plurality of three-dimensional target simulation models according to the point cloud fitting results; finally, building an anomaly recognition plug-in based on a predetermined monitoring index and the operation warning area, and using the anomaly recognition plug-in to perform anomaly detection on the plurality of three-dimensional target simulation models, and outputting a plurality of operation detection results; it can realize the comprehensive monitoring and anomaly detection of high-altitude electric power operations, significantly improve the timeliness, comprehensiveness and accuracy of operation safety monitoring, so as to timely discover potential safety hazards, effectively improve the safety, efficiency and accuracy during the operation process, and ensure the life safety of operation personnel and the quality of high-altitude operations.

[0106] Embodiment 2

[0107] As Figure 2 shown, based on the same inventive concept as the method for identifying and detecting the supervision and inspection target of an unmanned aerial vehicle provided in Embodiment 1, the embodiment of the present invention further provides a system for identifying and detecting the supervision and inspection target of an unmanned aerial vehicle, including:

[0108] The operation risk identification module 01 is used to identify the power operation risks based on the environmental characteristics and task attributes of the power operation area, and determine the operation risk coefficient and the operation warning area; the operation data acquisition module 02 is used to configure the monitoring frequency according to the operation risk coefficient, and based on the predetermined monitoring plan, control the drone to collect data according to the monitoring frequency, and obtain the monitoring images and the monitoring point cloud data set, wherein the monitoring frequency is positively correlated with the operation risk coefficient; the monitoring image segmentation module 03 is used to construct a target area segmenter based on the predetermined operation monitoring target, and use the target area segmenter to segment the monitoring image to output the area image set; the simulation model construction module 04 is used to traverse and compare the area image set with the historical area image set of the previous adjacent monitoring node, select the area images with the similarity less than the preset similarity threshold as the significantly different images, and based on the significantly different image set, perform point cloud fitting in the monitoring point cloud data set respectively, and construct multiple three-dimensional target simulation models according to the point cloud fitting results; the target anomaly detection module 05 is used to build an anomaly recognition plug-in based on the predetermined monitoring index and the operation warning area, and use the anomaly recognition plug-in to detect the anomalies of the multiple three-dimensional target simulation models, and output multiple operation detection results.

[0109] Further, a drone safety supervision and inspection target recognition and detection system further includes: reading the environmental characteristics and task attributes of the power operation area, wherein the environmental characteristics at least include wind force, wind direction, brightness and temperature, and the task attributes at least include task type and operation height; taking the area characteristics of the power operation area as a constraint, retrieving and obtaining the sample environmental characteristic set and the sample task attribute set, and marking the operation risk coefficient and the operation warning area under different sample environmental characteristics and different sample task attributes by domain experts to obtain the sample risk coefficient set and the sample warning area set; using the sample environmental characteristics and the sample task attributes as inputs and the sample risk coefficient and the sample warning area as supervision, using the sample environmental characteristic set, the sample task attribute set, the sample risk coefficient set and the sample warning area set to supervise and train the BP neural network until convergence to obtain the operation risk identification channel; inputting the environmental characteristics and the task attributes into the operation risk identification channel, and outputting the operation risk coefficient and the operation warning area.

[0110] Further, a drone safety supervision and inspection target recognition and detection system further includes: obtaining the predetermined operation monitoring target, wherein the predetermined operation monitoring target at least includes the operation position, the operation component, the operation tool and the operation personnel; taking the area characteristics, the environmental characteristics and the task attributes as retrieval constraints, collecting multiple sample operation images, and marking and selecting different sample operation images according to the predetermined operation monitoring target to obtain multiple sample area image sets; using the multiple sample operation images and the multiple sample area image sets as training data to supervise and train the convolutional neural network until convergence to generate the target area segmenter; inputting the monitoring image into the target area segmenter, and outputting the area image set.

[0111] Furthermore, a UAV safety supervision and inspection target recognition and detection system further includes: taking regional features, environmental features, task attributes, and a predetermined operation monitoring target as retrieval constraints, collecting multiple groups of sample region images, and obtaining multiple sample image similarities through identification, where each group of region images consists of two images containing the same monitoring target; training a Siamese network using the multiple groups of sample region images and the multiple sample image similarities to obtain a converged power image comparison model; using the power image comparison model to traverse and compare the region image set with the historical region image set of the previous adjacent monitoring node, and outputting multiple image similarities; screening the multiple image similarities according to a preset similarity threshold, and setting the region images with similarities less than the preset similarity threshold as significantly different images, and outputting a significantly different image set.

[0112] Furthermore, a UAV safety supervision and inspection target recognition and detection system further includes: randomly selecting a first significantly different image from the significantly different image set, and extracting contour features of the first significantly different image to output first contour features; performing point cloud random fitting in the monitoring point cloud dataset based on the first contour features to obtain a first point cloud fitting result, and calculating a first fitting degree, where the first fitting degree is the ratio of the number of point clouds falling within the first contour features to the total number of point clouds in the monitoring point cloud dataset; performing point cloud random fitting again in the monitoring point cloud dataset according to the first contour features to obtain a second point cloud fitting result and a second fitting degree; continuing iterative fitting until a predetermined number of times is reached, outputting multiple point cloud fitting results and multiple fitting degrees, and selecting the point cloud fitting result with the maximum fitting degree as the first optimal point cloud fitting result; successively obtaining multiple optimal point cloud fitting results of multiple significantly different images in the significantly different image set, and simulating and constructing multiple three-dimensional target simulation models according to the multiple optimal point cloud fitting results.

[0113] Further, a UAV safety supervision and inspection target recognition and detection system further includes: configuring predetermined monitoring indicators, where the predetermined monitoring indicators correspond to predetermined operation monitoring targets and include position monitoring indicators, component monitoring indicators, tool monitoring indicators, and personnel safety indicators; randomly selecting a first monitoring target from the predetermined operation monitoring targets and obtaining the first monitoring indicators of the first monitoring target; retrieving and obtaining multiple first sample three-dimensional target models based on the first monitoring target, and respectively performing normal / abnormal labeling on the multiple first sample three-dimensional target models according to the first monitoring indicators and the operation warning area to obtain multiple first sample labeling results; using the multiple first sample three-dimensional target models and the multiple first sample labeling results as training data, equally dividing the training data into Q parts, selecting Q times with replacement, constructing a first training set, iteratively selecting Q times, and obtaining Q training sets, where Q is an integer greater than 5; using the Q training sets to perform supervised training on a convolutional neural network to obtain Q first anomaly recognition branches, and constructing a first anomaly recognition unit based on the Q first anomaly recognition branches; sequentially obtaining multiple anomaly recognition units for multiple monitoring targets, and based on the mapping association between the monitoring targets and the anomaly recognition units, fusing and constructing an anomaly recognition plug-in according to the multiple anomaly recognition units of the multiple monitoring targets.

[0114] Further, a UAV safety supervision and inspection target recognition and detection system further includes: respectively calculating the similarity deviations between multiple image similarities and a preset similarity threshold to obtain multiple similarity deviations; respectively calculating the ratios of the multiple similarity deviations to the multiple historical maximum similarity deviations to determine multiple deviation scale coefficients, where the deviation scale coefficients correspond to the monitoring targets one by one; multiplying the multiple deviation scale coefficients by Q respectively to obtain multiple branch call numbers; based on the multiple branch call numbers, randomly selecting the same number of anomaly recognition branches in the multiple anomaly recognition units of the anomaly recognition plug-in to generate an adapted anomaly recognition plug-in; using the adapted anomaly recognition plug-in to perform anomaly detection on multiple three-dimensional target simulation models and outputting multiple operation detection results for multiple monitoring targets.

[0115] Embodiment 3

[0116] Please refer to Figure 3 , Figure 3 which is a schematic diagram of an embodiment of the electronic device provided by the embodiment of the present invention. As Figure 3As shown in the figure, an embodiment of the present invention provides an electronic device 500, including a memory 510, a processor 520, and a first computer program 511 stored on the memory 510 and executable on the processor 520. When the processor 520 executes the first computer program 511, the following steps are implemented: performing power operation risk identification based on the environmental characteristics and task attributes of the power operation area, determining the operation risk coefficient and the operation warning area; configuring the monitoring frequency according to the operation risk coefficient, and based on a predetermined monitoring scheme, controlling the drone to perform data collection according to the monitoring frequency, obtaining monitoring images and a monitoring point cloud data set, wherein the monitoring frequency is positively correlated with the operation risk coefficient; constructing a target area segmenter based on a predetermined operation monitoring target, using the target area segmenter to perform target segmentation on the monitoring images, and outputting a set of area images; traversing and comparing the set of area images with the historical set of area images of the previous adjacent monitoring node, selecting the area images with a similarity less than a preset similarity threshold as significantly different images, and based on the set of significantly different images, performing point cloud fitting in the monitoring point cloud data set respectively, and constructing a plurality of three-dimensional target simulation models according to the point cloud fitting results; building an anomaly recognition plug-in based on a predetermined monitoring index and the operation warning area, and using the anomaly recognition plug-in to perform anomaly detection on the plurality of three-dimensional target simulation models, and outputting a plurality of operation detection results.

[0117] Embodiment 4

[0118] Please refer to Figure 4 , Figure 4 which is a schematic diagram of an embodiment of a computer-readable storage medium provided by an embodiment of the present invention. As Figure 4 shown, this embodiment provides a computer-readable storage medium 600, on which a second computer program 611 is stored. When the second computer program 611 is executed by a processor, the following steps are implemented: performing power operation risk identification based on the environmental characteristics and task attributes of the power operation area, determining the operation risk coefficient and the operation warning area; configuring the monitoring frequency according to the operation risk coefficient, and based on a predetermined monitoring scheme, controlling the drone to perform data collection according to the monitoring frequency, obtaining monitoring images and a monitoring point cloud data set, wherein the monitoring frequency is positively correlated with the operation risk coefficient; constructing a target area segmenter based on a predetermined operation monitoring target, using the target area segmenter to perform target segmentation on the monitoring images, and outputting a set of area images; traversing and comparing the set of area images with the historical set of area images of the previous adjacent monitoring node, selecting the area images with a similarity less than a preset similarity threshold as significantly different images, and based on the set of significantly different images, performing point cloud fitting in the monitoring point cloud data set respectively, and constructing a plurality of three-dimensional target simulation models according to the point cloud fitting results; building an anomaly recognition plug-in based on a predetermined monitoring index and the operation warning area, and using the anomaly recognition plug-in to perform anomaly detection on the plurality of three-dimensional target simulation models, and outputting a plurality of operation detection results.

[0119] It should be noted that in the above embodiments, the descriptions of each embodiment have their own emphases. For parts not described in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0120] Those skilled in the art should understand that the embodiments of the present invention can be provided as a method, a system, or a computer program product. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0121] The present invention is described with reference to the flowcharts and / or block diagrams of methods, apparatuses (systems), and computer program products according to the embodiments of the present invention. It should be understood that each flow and / or block in the flowchart and / or block diagram, as well as the combination of flows and / or blocks in the flowchart and / or block diagram, can be realized by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded computer, or other programmable data processing devices to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing devices generate a device for realizing the functions specified in Figure 1 one flow or multiple flows and / or blocks Figure 1 one block or multiple blocks.

[0122] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer-readable memory generate a manufactured article including an instruction device, and the instruction device realizes the functions specified in Figure 1 one flow or multiple flows and / or blocks Figure 1 one block or multiple blocks.

[0123] These computer program instructions can also be loaded onto a computer or other programmable data processing device, so that a series of operation steps are executed on the computer or other programmable device to generate a computer-implemented process, and thus the instructions executed on the computer or other programmable device provide steps for realizing the functions specified in Figure 1 one flow or multiple flows and / or blocks Figure 1 one block or multiple blocks.

[0124] The above are only the preferred embodiments of the present invention, and do not impose any form of limitation on the present invention. Any person skilled in the relevant art, without departing from the scope of the technical solution of the present invention and based on the technical essence of the present invention, any simple modifications, equivalent replacements, improvements, etc. made to the above embodiments shall still fall within the scope of protection of the technical solution of the present invention.

Claims

1. A method for identifying and detecting supervision and inspection targets of an unmanned aerial vehicle for power operation, characterized in that, Including the following steps: S1: Identify the power operation risks based on the environmental characteristics and task attributes of the power operation area, and determine the operation risk coefficient and the operation warning area; S2: Configure the monitoring frequency according to the operation risk coefficient, and based on the predetermined monitoring plan, control the drone to collect data from the power operation area at the monitoring frequency to obtain monitoring images and a monitoring point cloud dataset, where the monitoring frequency is positively correlated with the operation risk coefficient; S3: Set the monitoring time node according to the actual scenario, and build a target area segmenter based on the predetermined operation monitoring target. Use the target area segmenter to perform target segmentation on the monitoring images and output a set of area images; S4: Traverse and compare the set of area images collected at the current monitoring node with the historical set of area images collected at the previous adjacent monitoring node. Select the area images with a similarity less than the preset similarity threshold as significantly different images, and based on the obtained set of significantly different images, perform point cloud fitting in the monitoring point cloud dataset respectively. Build a three-dimensional target simulation model according to the point cloud fitting results; S5: Build an anomaly recognition plugin based on the predetermined monitoring indicators and the operation warning area, and use the anomaly recognition plugin to perform anomaly detection on the three-dimensional target simulation model and output the operation detection result.

2. The power operation UAV safety supervision and inspection target recognition and detection method according to claim 1, characterized in that: Step S1 includes: Taking the area characteristics of the power operation area as a constraint, retrieve and obtain a sample environmental characteristic set and a sample task attribute set, and have domain experts label the operation risk coefficient and the operation warning area for them to obtain a sample risk coefficient set and a sample warning area set; Using the sample environmental characteristics and sample task attributes as inputs, and the sample risk coefficient set and sample warning area set as supervision, perform supervised training with a BP neural network until convergence to obtain an operation risk recognition channel for identifying and determining the operation risk coefficient and the operation warning area.

3. The power operation UAV safety supervision and inspection target recognition and detection method according to claim 1, characterized in that: Step S3 includes: Obtain the predetermined operation monitoring target, take the area characteristics, environmental characteristics, and task attributes as retrieval constraints, collect sample operation images, and perform annotation and bounding on different sample operation images according to the predetermined operation monitoring target to obtain a sample set of area images; Using the sample operation images and the sample set of area images as training data, perform supervised training on a convolutional neural network until convergence to generate a target area segmenter for performing target segmentation on the monitoring images.

4. The power operation UAV safety supervision and inspection target recognition and detection method according to claim 3, characterized in that: Step S4 includes: Taking the area characteristics, environmental characteristics, task attributes, and predetermined operation monitoring target as retrieval constraints, collect a sample set of area images. Each set of area images in the sample set of area images is two images containing the same monitoring target, and perform similarity annotation on the two images containing the same monitoring target; Use the sample set of area images and the marked similarity to train a siamese network to obtain a power image comparison model that meets convergence; Use the power image comparison model to traverse and compare the set of area images with the historical set of area images at the previous adjacent monitoring node, and output the image similarity; Screen the obtained image similarity according to the preset similarity threshold, select the area images with a similarity less than the preset similarity threshold as significantly different images, and output the set of significantly different images.

5. The power operation UAV safety supervision and inspection target recognition and detection method according to claim 4, characterized in that: Step S4 further includes: Randomly select a first significantly different image from the significantly different image set, extract the contour features of the first significantly different image, and output the first contour features; Based on the first contour features, perform point cloud random fitting on the monitored point cloud dataset to obtain a first point cloud fitting result, and calculate a first fitting degree; wherein, the first fitting degree is the ratio of the number of point clouds falling within the first contour features to the total number of point clouds in the monitored point cloud dataset; Based on the first contour features, perform the next point cloud fitting on the monitored point cloud dataset to obtain a second point cloud fitting result and a second fitting degree; Continue iterative fitting until a predetermined number of times is reached, output multiple point cloud fitting results and multiple fitting degrees, and select the point cloud fitting result with the largest fitting degree as the first optimal point cloud fitting result; Successively obtain the optimal point cloud fitting results of all significantly different images in the significantly different image set, and respectively simulate and construct a three-dimensional target simulation model according to the respective optimal point cloud fitting results.

6. The power operation UAV safety supervision and inspection target recognition and detection method according to claim 5, characterized in that: Step S5 includes: Configure a predetermined monitoring index corresponding to the predetermined operation monitoring target; Randomly select a first monitoring target from the predetermined operation monitoring targets, and obtain the first monitoring index of the first monitoring target; Based on the first monitoring target, retrieve and obtain a first sample three-dimensional target model from the historical three-dimensional model database, and perform normal / abnormal annotation on the first sample three-dimensional target model according to the first monitoring index and the operation warning area to obtain a first sample annotation result; Use the first sample three-dimensional target model and the first sample annotation result as training data, randomly divide the training data into Q parts, construct Q training sets through sampling with replacement, construct a first training set, and iteratively select Q times to obtain Q training sets; Use the Q training sets to perform supervised training on the convolutional neural network to obtain Q first anomaly recognition branches, and construct a first anomaly recognition unit based on the Q first anomaly recognition branches; Successively obtain the anomaly recognition units corresponding to each monitoring target in the predetermined operation monitoring target, and fuse and construct an anomaly recognition plugin based on the mapping association relationship between the monitoring target and the anomaly recognition unit.

7. The electric power operation UAV safety supervision and inspection target recognition and detection method according to claim 6, characterized in that: The method for fusing and constructing an anomaly recognition plugin based on the mapping association relationship between the monitoring target and the anomaly recognition unit is: Calculate the similarity deviation between each image similarity and a preset similarity threshold respectively; Calculate the ratio of the similarity deviation to the corresponding historical maximum similarity deviation respectively to determine each deviation scale coefficient, wherein each deviation scale coefficient corresponds to each monitoring target one by one; Multiply each deviation scale coefficient by Q to obtain a plurality of branch call numbers; Based on the branch call numbers, randomly select the same number of anomaly recognition branches from the multiple anomaly recognition units of the anomaly recognition plugin to generate an adapted anomaly recognition plugin; Use the adapted anomaly recognition plugin to perform anomaly detection on the three-dimensional target simulation model and output the operation detection result of the monitoring target.

8. An inspection target recognition and detection system for power operation unmanned aerial vehicles, characterized in that: The method for identifying and detecting the safety supervision target of an electric power operation unmanned aerial vehicle for implementing any one of claims 1 to 7 includes: The operation risk identification module is used to identify the risks of electric power operations based on the environmental characteristics and task attributes of the electric power operation area, and determine the operation risk coefficient and the operation warning area; The operation data acquisition module is used to configure the monitoring frequency according to the operation risk coefficient, and based on a predetermined monitoring scheme, control the drone to perform data acquisition according to the monitoring frequency, and obtain the monitoring images and the monitoring point cloud data set; The monitoring image segmentation module is used to construct a target area segmenter based on a predetermined operation monitoring target, and use the target area segmenter to perform target segmentation on the monitoring image, and output a set of area images; The simulation model construction module is used to traverse and compare the set of area images with the historical set of area images of the previous adjacent monitoring node, select the area images with a similarity less than a preset similarity threshold as significantly different images, and based on the set of significantly different images, perform point cloud fitting in the monitoring point cloud data set respectively, and construct a three-dimensional target simulation model according to the point cloud fitting result; The target anomaly detection module is used to build an anomaly recognition plug-in based on a predetermined monitoring index and the operation warning area, and use the anomaly recognition plug-in to perform anomaly detection on the three-dimensional target simulation model, and output an operation detection result.

9. An electronic device, characterized in that, Including: A memory, a processor, and a computer program stored on the memory and executable on the processor, wherein when the processor executes the computer program, the method for identifying and detecting the safety supervision and inspection target of an electric power operation drone as described in any one of claims 1 to 7 is implemented.

10. A computer-readable storage medium storing a computer program, wherein the computer program causes a computer to execute the method for identifying and detecting the safety supervision and inspection target of an electric power operation drone as described in any one of claims 1 to 7.

Images