A method for strengthening the control of distribution network projects

Abstract

The present invention discloses a method for strengthening the control of distribution network projects, and the steps are as follows: S1. Set the line s and speed v of the supervision robot; S2. Shoot the supervision video according to the line s and speed v, and record the motion equation F of the camera; S3. Annotate the collected video to produce a standard scene image library; S4. Produce a target data set for the distribution network construction site; S5. Train the YOLOV5s deep neural network to obtain training weights; S6. Produce negative samples of the standard scene image library; S7. Train the MatchNet neural network based on probability similarity on the positive and negative samples; S8. The supervision robot conducts supervision according to the line x, speed v and camera motion equation F; S9. Compare the captured images with the pictures in the standard scene library by using the MatchNet neural network based on probability similarity; S10. Use the trained YOLOV5s deep neural network to detect abnormalities in the abnormal pictures; S11. Output the results of the comparison abnormalities. This solution improves the accuracy and efficiency of abnormal monitoring.

Description

Technical Field

[0001] The present invention relates to the technical field of power engineering management and control, and specifically, to a method for strengthening the management and control of distribution network projects. Background Art

[0002] With the rapid development of social economy, the demand for electric power resources by people is increasing. At the same time, distribution network projects are an important part of the power industry and the infrastructure work of the company. However, there are problems with backward safety management and control technologies in existing distribution network project operations, and it is difficult to detect potential safety hazards in a timely manner. Therefore, it is urgent to improve the strengthening management and control technology of distribution network projects.

[0003] In recent years, the application of computer vision technology based on deep learning in the strengthening management and control of distribution networks has also been increasing, such as performing anomaly detection and early warning processing for pole tilt, insufficient guy wire angle, insufficient equipment installation height, etc. T. Würf et al. proposed a new deep learning framework for 3D tomography reconstruction for the problem of 3D image recognition, mapping the filtered back-projection type algorithm to a neural network, and accurately projecting a 3D image into a 2D image for recognition. L. Chen et al. used atrous convolution to explicitly control the calculation of feature responses in a CNN, proposed the ASPP algorithm to robustly segment images at multiple scales, combined DCNNs and probabilistic graphical models, thereby improving the method for locating the boundary of the target image and increasing the accuracy of image meaning recognition. Qiao Junfeng et al. studied the compatibility between big data technology and distribution network monitoring by analyzing the implementation elements of panoramic monitoring of the distribution network, and proposed an implementation method for panoramic monitoring of the distribution network based on big data technology. Li Junfeng et al. extracted the image features of power equipment through a CNN, borrowed traditional machine learning methods, and proposed a random forest classification method combined with deep learning. A large number of power equipment images were used to test the model, and the results showed that the average recognition accuracy for power equipment such as insulators, transformers, and circuit breakers was significantly improved.

[0004] The above research results are of great significance for strengthening the management and control of distribution network projects. However, the existing technologies still have certain limitations. First, the above methods do not achieve systematization, and are only the detection of on-site targets in the distribution network by image processing technology. Second, the deep learning neural network algorithms used in the above methods have very high requirements for device computing power and are difficult to deploy on embedded devices. Finally, the above methods are only applicable to local targets on the distribution network site and cannot perform real-time anomaly monitoring on the entire distribution network site, with a low degree of intelligence. Summary of the Invention

[0005] The objective of the present invention is to solve the problem that the existing engineering management methods have a low level of intelligence and cannot monitor the distribution network site in the full time domain. A method for strengthening the control of distribution network projects is proposed. Based on the object detection technology and image matching technology of deep learning, a distribution network supervision robot is used to realize the real-time control of the quality and safety of the distribution network site, reducing the hidden dangers of traditional manual control and improving the accuracy and efficiency of abnormal monitoring.

[0006] To achieve the above technical objectives, a technical solution provided by the present invention is a method for strengthening the control of distribution network projects, including the following steps:

[0007] Step S1: Set the supervision route s and supervision speed v of the supervision robot at the distribution network construction site;

[0008] Step S2: The supervision robot equipped with an RGB camera conducts supervision according to the route s and speed v set in Step S1, shoots the supervision video, and records the motion equation F of the camera;

[0009] Step S3: Mark the types, quantities, and positions of the distribution network site targets involved in each frame of the collected video to form a standard scene image library;

[0010] Step S4: Collect the distribution network site targets involved in Step S3 and make them into a distribution network construction site target data set;

[0011] Step S5: Train the YOLOV5s deep neural network on the data set in Step S4 to obtain the training weights;

[0012] Step S6: Make negative samples of the standard scene image library in Step S3;

[0013] Step S7: Train the MatchNet neural network based on probability similarity on the positive and negative samples to obtain the final network;

[0014] Step S8: The supervision robot conducts supervision according to the route and speed set in Step S1 and the camera motion equation F recorded in Step S2;

[0015] Step S9: During the supervision process, compare each frame of the captured image with 30 pictures before and after the same moment in the standard scene library in Step S3 using the MatchNet neural network based on probability similarity;

[0016] Step S10: Use the trained YOLOV5s deep neural network in Step S5 to perform final abnormal detection on the pictures preliminarily compared as abnormal in Step S9;

[0017] Step S11: Output the results of the comparison as abnormal.

[0018] In this solution, combining the object detection technology and image matching technology of deep learning, a distribution network supervision robot is used to achieve real-time control of the quality and safety of the distribution network site. First, according to the fixed supervision route and speed, the camera is rotated to collect the video of the distribution network site, and the motion equation of the camera is recorded; each frame of the collected video is parsed into pictures, and all the targets involved in the distribution network site are marked, including the types and quantities of the targets; a dataset containing the distribution network site targets in the collected video is made, which can be obtained through the network or shot on the spot, and marked to form training samples in YOLO format as the standard scene image library to train the YOLOV5s object detection network; at the same time, negative samples of the standard scene image library are made, and positive and negative samples are used to train the MatchNet object matching neural network. In order to realize the fusion of YOLOV5s and MatchNet, a MatchNet algorithm based on probability similarity is designed, which can output specific probability values instead of 0 and 1 values; a similarity threshold is set, and the scenes detected as abnormal by MatchNet are further confirmed by YOLOV5s and the specific abnormal parts are detected; finally, the abnormal logs are saved for information-based control. This method is not limited by the distribution network site scene, has good generalization performance for the distribution network sites of different power grids, does not need to train models separately for each different distribution network site additionally, reduces some repetitive image annotation and model training work, and is convenient for development and deployment. At the same time, it helps developers and power grid managers improve the efficiency of warning about potential safety hazards of the staff at the distribution network site.

[0019] Preferably, step S3 includes the following steps:

[0020] Parse the collected supervision video into each frame, and use the labelme annotation tool to mark the distribution network construction site targets in each picture to generate an annotation file conforming to the YOLOV5 neural network;

[0021] The annotation information of each picture includes the target type, target quantity and position information in the picture; it is expressed as follows:

[0022] class_id,x,y,w,h

[0023] class_id represents the target category number, which increases sequentially starting from 0; (x,y) is the ratio of the coordinates of the center of the true box relative to the upper left corner of the picture to the width and height of the picture; (w,h) is the ratio of the width and height of the true box to the width and height of the picture;

[0024] All the annotated files and supervision videos after annotation are used as the standard scene library.

[0025] Preferably, step S5 includes the following steps:

[0026] Cluster 9 prior box sizes on the dataset in step S4 using the kmeans++ algorithm;

[0027] The total number of training rounds of the YOLOV5s algorithm is 500, and the initial learning rate is set to 0.0013. Let it be reduced to 1 / 10 of the original value at 400,000 rounds and 450,000 rounds of training respectively, with decay set to 0.0005. During the training process, rotate the pictures and change the hue and saturation to prevent overfitting;

[0028] The class loss function uses the cross-entropy loss function, as shown in the following formula:

[0029]

[0030] Among them, L class represents the class loss of the prediction box; represents the actual probability that when the i-th grid detects a target, this target belongs to class c; represents the predicted probability that when the i-th grid detects a target, this target belongs to class c; c represents the class to which the detected target belongs;

[0031] The bounding box regression loss function uses the GIoU loss function, and the principle formula of the function is as follows:

[0032]

[0033] L GIoU = 1 - GIoU

[0034] In the above formula, L GIoU represents the GIoU loss; P represents the prediction box; G represents the ground truth box; C represents the area of the smallest bounding rectangle that contains both the prediction box and the ground truth box;

[0035] The total loss function is:

[0036] L loss = L class + L GIoU .

[0037] Preferably, step S6 includes the following steps:

[0038] Randomly perform mask occlusion, brightness, grayscale change, and blurring processing on each picture in the standard scene image library in step S3 to obtain a negative sample image library;

[0039] Divide each picture in the standard scene image library and the negative sample image library into 64×64 image blocks to obtain positive and negative samples for training.

[0040] Preferably, step S7 includes the following steps:

[0041] Build a feature extraction network based on Keras and Python languages; the feature extraction network consists of two identical neural network structures. A single neural network structure is as follows: a convolutional layer of 7×7×24, a max pooling layer of 3×3 / 2, a convolutional layer of 5×5×64, a max pooling layer of 3×3 / 2, two convolutional layers of 3×3×96, a convolutional layer of 3×3×64, and a max pooling layer of 3×3 / 2. The stride of the convolutional kernel in all convolutional layers is 1, and the stride of the max pooling layer is 2.

[0042] Build a feature matching network based on Keras and Python languages; the feature matching network consists of three fully connected layers plus a softmax layer. Each of the two fully connected layers has 1024 neurons, the activation function is relu, and the softmax layer has 1000 neurons, and the activation function is softmax.

[0043] Preferably, step S9 includes the following steps:

[0044] Considering the internal assembly error of the supervision robot and the environmental changes in different supervision periods, each captured image is compared with 30 images before and after the same moment. Select one image from the standard scene library with the highest output similarity probability. If there are multiple images with the same probability, select the image with a smaller time as the comparison object;

[0045] Since the MatchNet neural network based on probability similarity divides the image into sub-blocks of 64×64, it is set that when no more than 16 sub-blocks are different, the two images are considered to be exactly the same. Therefore, the similarity threshold is:

[0046]

[0047] That is, an image scene with a similarity greater than or equal to this threshold p can be considered normal; when the similarity is less than this threshold p, the scene can be considered preliminarily abnormal.

[0048] Preferably, step S10 includes the following steps:

[0049] Perform object detection on the two images with abnormal comparison using the YOLOV5s deep neural network to obtain the probability and detection frame of each object;

[0050] Use a thick line box to mark the object whose target detection probability, detection frame width, height, or width-to-height ratio in the typical scene library differs by more than 0.004. This object is the final abnormal object;

[0051] If the difference in probability or detection frame parameters is less than 0.004, it is considered normal.

[0052] Preferably, step S11 includes the following steps: saving the abnormal pictures, the abnormal targets, and the moments of the abnormal pictures, and outputting them to a log document.

[0053] Advantages of the present invention: A method for strengthening the control of a power distribution project according to the present invention can achieve all-weather real-time safety control of the power distribution site based on deep learning computer vision technology, reducing the hidden dangers of traditional manual control; the data set of the power distribution site is general, and the power distribution supervision robot equipped with the power distribution safety control technology of the present invention can be applied to any power distribution site. The fusion of YOLOV5s and the MatchNet algorithm improves the accuracy of scene matching on the basis of object detection, and improves the accuracy and efficiency of abnormal monitoring of the power distribution site. Description of the Drawings

[0054] Figure 1 It is a flowchart of a method for strengthening the control of a power distribution project according to the present invention. Detailed Embodiments

[0055] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Apparently, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0056] The terms "first", "second", "third", "fourth", etc. (if any) in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that such data may be interchanged under appropriate circumstances so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein.

[0057] It should be understood that in various embodiments of the present invention, the magnitude of the sequence numbers of the various processes does not mean the order of execution, and the order of execution of the various processes should be determined by their functions and internal logics, and should not constitute any limitation to the implementation process of the embodiments of the present invention.

[0058] It should be understood that in the present invention, "including" and "having" and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units does not necessarily have to be limited to those steps or units clearly listed, but may include other steps or units not clearly listed or inherent to these processes, methods, products, or devices.

[0059] It should be understood that in the present invention, "a plurality of" means two or more. "And / or" is merely a description of the relationship between associated objects, indicating that there can be three relationships. For example, A and / or B can represent: A exists alone, A and B exist simultaneously, and B exists alone. The character " / " generally indicates that the associated objects before and after are in an "or" relationship. "Including A, B, and C" and "including A, B, C" mean that all of A, B, and C are included. "Including A, B, or C" means that one of A, B, and C is included. "Including A, B, and / or C" means including any one or any two or all three of A, B, and C.

[0060] It should be understood that in the present invention, "B corresponding to A", "B corresponding to A", "A corresponding to B", or "B corresponding to A" means that B is associated with A, and B can be determined according to A. Determining B according to A does not mean determining B only according to A. B can also be determined according to A and / or other information. The matching of A and B means that the similarity between A and B is greater than or equal to a preset threshold.

[0061] Depending on the context, as used herein, "if" can be interpreted as "when", "while", "in response to determination", or "in response to detection".

[0062] The technical solution of the present invention will be described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments.

[0063] Embodiment:

[0064] As Figure 1 shown, a flowchart of a method for strengthening the control of a distribution network project includes the following steps:

[0065] Step S1: Set the supervision route s and supervision speed v of the supervision robot at the distribution network construction site; the purpose of fixing the supervision route and supervision speed is to reduce the difficulty and time of subsequent scene matching and improve the accuracy of scene matching.

[0066] Step S2: The supervision robot equipped with an RGB camera conducts supervision according to the route s and speed v set in step S1, shoots a supervision video, and records the motion equation F of the camera; when the supervision robot conducts supervision for the first time, it rotates the camera to shoot the supervision video to ensure that as many targets at the distribution network construction site as possible are shot on the supervision route set in step S1; the state of the camera rotation is recorded in real time, and when the supervision process is completed, the motion equation F of the camera is generated.

[0067] Step S3: Mark the types, quantities, and positions of the distribution network site targets involved in each frame of the collected video to form a standard scene image library; it includes the following steps:

[0068] Parse the collected supervision video into each frame, and use the labelme annotation tool to annotate the distribution network construction site targets in each picture to generate an annotation file that conforms to the YOLOV5 neural network;

[0069] The annotation information of each picture includes the target type, the number of targets, and the position information in the picture; it is expressed as follows:

[0070] class_id,x,y,w,h

[0071] class_id represents the target category number, which increases sequentially starting from 0; (x,y) is the ratio of the coordinates of the center of the true box relative to the upper left corner of the picture to the width and height of the picture; (w,h) is the ratio of the width and height of the true box to the width and height of the picture;

[0072] All the annotated files and the supervision video after annotation are used as the standard scene library.

[0073] Step S4: Collect the distribution network site targets involved in step S3 and make them into a distribution network construction site target dataset; download the target pictures involved in step S3 from the network or take pictures on the spot, and use the annotation method in step S3 for annotation to form a distribution network construction site target dataset, and divide it into a training set and a validation set according to a ratio of 7:3.

[0074] Step S5: Train the YOLOV5s deep neural network on the dataset in step S4 to obtain training weights; the training environment operating system is Ubuntu16.04; the CPU is Intel i7-8700, 3.2GHz; the memory size is 16G; the GPU is NVIDIA 1070 8GB graphics card; the deep learning framework is AlexeyAB-Darknet, and the source code is in C language;

[0075] Use the kmeans++ algorithm to cluster 9 types of prior box sizes on the dataset in step S4;

[0076] The total number of training rounds of the YOLOV5s algorithm is 500500, the initial learning rate is set to 0.0013, and it is reduced to 1 / 10 of the original value at 400000 rounds and 450000 rounds of training respectively, the decay is set to 0.0005. During the training process, the pictures are rotated and the hue and saturation are changed to prevent overfitting;

[0077] The category loss function uses the cross-entropy loss function, as shown in the following formula:

[0078]

[0079] Among them, L class represents the category loss of the prediction box; Represents the actual probability that when the i-th grid detects a target, this target belongs to class c; Represents the predicted probability that when the i-th grid detects a target, this target belongs to class c; c represents the class to which the detected target belongs;

[0080] The bounding box regression loss function adopts the GIoU loss function, and the functional principle formula is as follows:

[0081]

[0082] L GIoU = 1 - GIoU

[0083] In the above formula, L GIoU Represents the GIoU loss; P represents the predicted box; G represents the ground truth box; C represents the area of the smallest bounding rectangle that contains both the predicted box and the ground truth box;

[0084] The total loss function is:

[0085] L loss = L class + L GIoU .

[0086] Step S6, create negative samples for the standard scene image library in step S3; randomly perform mask occlusion, brightness and darkness, grayscale change, and blurring processing on each picture in the standard scene image library in step S3 to obtain a negative sample image library;

[0087] Divide each picture in the standard scene image library and the negative sample image library into 64×64 image patches to obtain positive and negative samples for training.

[0088] Step S7, train the MatchNet neural network based on probability similarity on the positive and negative samples to obtain the final network; including the following steps:

[0089] Construct a feature extraction network based on Keras and Python languages; the feature extraction network is composed of two identical neural network structures, and a single neural network structure is: a convolutional layer of 7×7×24, a max pooling layer of 3×3 / 2, a convolutional layer of 5×5×64, a max pooling layer of 3×3 / 2, two convolutional layers of 3×3×96, a convolutional layer of 3×3×64, and a max pooling layer of 3×3 / 2, where the stride of the convolutional kernel in all convolutional layers is 1, and the stride of the max pooling layer is 2;

[0090] Build a feature matching network based on Keras and the Python language; the feature matching network consists of three fully connected layers plus a softmax layer. Each of the two fully connected layers has 1024 neurons, with the activation function being relu, and the softmax layer has 1000 neurons, with the activation function being softmax.

[0091] Step S8: The supervision robot conducts supervision according to the route and speed set in step S1 and the camera motion equation recorded in step S2; since the same route x, speed, and the same camera motion equation F as those for collecting the standard scene image library are adopted, theoretically, in the ideal situation where the scene has not changed at all, the single-frame pictures collected at the same moment should also be exactly the same. This is also the reason for setting the route, speed, and camera motion state equation in step S1.

[0092] Step S9: During the supervision process, each frame of the captured image is compared with 30 pictures before and after the same moment in the standard scene library in step S3 using the MatchNet neural network based on probability similarity; it includes the following steps:

[0093] Considering the internal assembly error of the supervision robot and the environmental changes in different supervision periods, the pictures collected each time are compared with 30 pictures before and after the same moment. One picture from the standard scene library with the highest output similarity probability is selected. If there are multiple pictures with the same probability, the picture with the smaller moment is selected as the comparison object;

[0094] Since the MatchNet neural network based on probability similarity divides the picture into 64×64 sub-blocks, therefore, when no more than 16 sub-blocks are different, the two pictures are considered exactly the same. Thus, the similarity threshold is:

[0095]

[0096] That is, the picture scene with a similarity greater than or equal to this threshold p can be considered normal; when the similarity is less than this threshold p, the scene can be considered preliminarily abnormal.

[0097] Step S10: The pictures that are preliminarily compared as abnormal in step S9 are subjected to final anomaly detection using the trained YOLOV5s deep neural network in step S5; it includes the following steps:

[0098] Perform object detection on the two pictures that are compared as abnormal using the YOLOV5s deep neural network to obtain the probabilities and detection frames of each object;

[0099] Use a thick line box to mark the objects whose target detection probability, detection frame width, height, or width-to-height ratio in the typical scene library differ by more than 0.004. This object is the final abnormal object;

[0100] If the probability or the difference in detection box parameters is less than 0.004, it is considered normal.

[0101] Step S11: Output the result of comparison anomaly. Save the abnormal pictures, the abnormal targets and the time of the abnormal pictures, and output them to the log document.

[0102] In this embodiment, combining the object detection technology and image matching technology of deep learning, a distribution network supervision robot is used to realize the real-time control of the quality and safety of the distribution network site. First, according to the fixed supervision route and speed, the camera is rotated simultaneously to collect the distribution network site video, and the motion equation of the camera is recorded; each frame of the collected video is parsed into pictures, and all the targets involved in the distribution network site are marked, including the types and quantities of the targets; a data set containing the distribution network site targets in the collected video is made, which can be obtained through the network or shot on the spot, and marked to form training samples in YOLO format as the standard scene image library to train the YOLOV5s object detection network; at the same time, negative samples of the standard scene image library are made, and positive and negative samples are used to train the MatchNet object matching neural network. In order to realize the fusion of YOLOV5s and MatchNet, a MatchNet algorithm based on probability similarity is designed, which can output specific probability values instead of 0 and 1 values; a similarity threshold is set, and the scenes detected as abnormal by MatchNet are further confirmed by YOLOV5s and the specific abnormal parts are detected; finally, the abnormal log is saved for information-based control. This method is not limited by the distribution network site scene, has good generalization performance for the distribution network sites of different power grids, does not need to train models separately for each different distribution network site additionally, reduces some repetitive image annotation and model training work, is convenient for development and deployment, and at the same time helps developers and power grid managers improve the efficiency of warning about potential safety hazards of the staff at the distribution network site.

[0103] The above specific implementation manners are the preferred implementation manners of a method for strengthening the control of a distribution network project of the present invention, and do not limit the specific implementation scope of the present invention. The scope of the present invention includes but is not limited to this specific implementation manner. All equivalent changes made according to the shape and structure of the present invention are within the protection scope of the present invention.

Claims

1. A method for strengthening the control of distribution network projects, characterized in that: It includes the following steps: Step S1: Establish a MatchNet neural network model based on probability similarity; Step S2: The supervision robot conducts supervision according to the set route, speed, and camera motion equation F; Step S3: During the supervision process, each frame of the captured image is compared with 30 pictures before and after the same moment in the standard scene library using the MatchNet neural network based on probability similarity; Step S4: The pictures with preliminary comparison anomalies in Step S3 are subjected to final anomaly detection using the YOLOV5s deep neural network; Step S5: Output the results of the comparison anomalies; Step S1 includes the following steps: Step S11: Set the supervision route s and supervision speed v of the supervision robot at the distribution network construction site; Step S12: The supervision robot equipped with an RGB camera conducts supervision according to the route s and speed v set in Step S11, captures the supervision video, and records the camera motion equation F; Step S13: Mark the type, quantity, and location of the distribution network site targets involved in each frame of the collected video to form a standard scene image library; Step S14: Collect the distribution network site targets involved in Step S13 and make them into a distribution network construction site target dataset; Step S15: Train the YOLOV5s deep neural network on the dataset in Step S14 to obtain the training weights; Step S16: Make negative samples of the standard scene image library in Step S13; Step S17: Train the MatchNet neural network based on probability similarity on the positive and negative samples to obtain the final network; Step S3 includes the following steps: Compare each captured picture with 30 pictures before and after the same moment, select one picture from the standard scene library with the highest output similarity probability. If there are multiple pictures with the same probability, select the picture with the smaller moment as the comparison object; The MatchNet neural network based on probability similarity divides the picture into 64×64 sub-blocks. It is set that when no more than 16 sub-blocks are different, the two pictures are considered completely consistent, and its similarity threshold is: ； That is, pictures with a scene greater than or equal to this threshold p can be considered normal; when the similarity is less than this threshold p, the scene can be considered preliminarily abnormal; Step S4 includes the following steps: Conduct target detection on the two pictures with comparison anomalies using the YOLOV5s deep neural network to obtain the probability of each target and the detection box; Use a thick line box to mark the targets whose target detection probability, detection box width, height, or width-to-height ratio in the typical scene library differ by more than 0.

004. This target is the final abnormal target; If the probability or detection box parameter difference is less than 0.004, it is considered normal; Use the kmeans++ algorithm to cluster 9 types of prior box sizes on the dataset in Step S14; During the training process, rotate the pictures and change the hue and saturation to prevent overfitting; The category loss function uses the cross-entropy loss function, as shown in the following formula: ； Among them, L class represents the class loss of the prediction box; represents when the i -th grid detects a target, the actual probability that this target belongs to class c ; represents when the i -th grid detects a target, the predicted probability that this target belongs to class c ; c represents the class to which the detected target belongs; The bounding box regression loss function uses the GIoU loss function, and the function principle formula is as follows: ； ； In the above formula L GIoU represents GIoU loss; P represents the predicted bounding box; G represents the ground truth bounding box; C represents the area of the smallest bounding rectangle that contains both the predicted bounding box and the ground truth bounding box; The total loss function is: 。 2. The enhanced control method for a distribution network project according to claim 1, characterized in that: Step S13 includes the following steps: Parse the collected supervision video into each frame, and use the labelme annotation tool to annotate the power distribution construction site targets in each picture, generating an annotation file that conforms to the YOLOV5s neural network; All the annotated files and the supervision video after annotation are used as the standard scene library.

3. The enhanced control method for a distribution network project according to claim 1 or 2, characterized in that: Step S16 includes the following steps: Randomly perform mask occlusion, brightness and darkness, gray-scale change, and blur processing on each picture in the standard scene image library in step S13 to obtain a negative sample image library; Divide each picture in the standard scene image library and the negative sample image library into 64×64 image blocks to obtain positive and negative samples for training.

4. A method for strengthening the control of a power distribution project according to claim 3, characterized in that: Step S17 includes the following steps: Construct a feature extraction network based on Keras and the Python language; the feature extraction network is composed of two identical neural network structures. A single neural network structure is: a convolutional layer of 7×7×24, a max-pooling layer of 3×3 / 2, a convolutional layer of 5×5×64, a max-pooling layer of 3×3 / 2, two convolutional layers of 3×3×96, a convolutional layer of 3×3×64, and a max-pooling layer of 3×3 / 2. Among them, the stride of the convolution kernel in all convolutional layers is 1, and the stride of the max-pooling layer is 2.

5. A method for strengthening the control of a power distribution project according to claim 4, characterized in that: Step S17 further includes the following steps: Build a feature matching network based on Keras and the Python language; the feature matching network is composed of three fully connected layers plus a softmax layer. Each of the two fully connected layers has 1024 neurons, the activation function is relu, and the softmax layer has 1000 neurons, and the activation function is softmax.

6. A method for strengthening the control of a power distribution project according to claim 1, characterized in that: Step S5 includes the following steps: Save the abnormal pictures, the abnormal targets, and the moments of the abnormal pictures, and output them to a log document.

7. A method for strengthening the control of a power distribution project according to claim 1, characterized in that: Step S14 includes the following steps: Download all the target pictures in step S13 from the network or take them on-site, and perform annotation using the annotation method in step S13 to form a power distribution construction site target data set, and divide it into a training set and a validation set according to a ratio of 7:3.

Images