Long-endurance unmanned aerial vehicle intelligent inspection method applied to power transmission and distribution lines

Abstract

The invention provides a long-endurance unmanned aerial vehicle intelligent inspection method applied to power transmission and distribution lines, and belongs to the technical field of unmanned aerial vehicle intelligent inspection control. The method comprises the following steps: 1) establishing an unmanned aerial vehicle intelligent inspection path planning model, and obtaining an unmanned aerial vehicle intelligent inspection optimal path based on MOPFA; 2) generating a data set by using the power transmission and distribution line images shot by the unmanned aerial vehicle, continuously training the faster RCNN model, and performing fault recognition; 3) constructing an orthogonal projector based on an orthogonal weight correction algorithm, and updating the fault recognition model in real time to improve the fault recognition precision; and 4) the identification result is fed back to the control center in real time, so that a dispatcher can formulate a corresponding coping scheme. According to the method, the inspection efficiency of the unmanned aerial vehicle can be greatly improved, and good performance is shown in fault identification.

Description

Technical Field

[0001] The present invention relates to the technical field of unmanned aerial vehicle (UAV) intelligent inspection, and in particular to a long-endurance unmanned aerial vehicle (UAV) intelligent inspection method for power transmission and distribution lines. Background Art

[0002] With the rapid development of the power system, the scale of my country's power grid is increasing. Among them, transmission and distribution lines play an important role in the transmission of electric energy and are the hubs connecting the power generation side and the user side. However, the distribution of transmission and distribution lines is relatively wide, mostly located in hilly, mountainous, watery and other areas. Due to the influence of the natural environment, hardware fittings may fall off, insulators may explode, and other faults may occur, which will have a huge impact on the transmission of electric energy. In order to grasp the operation of transmission and distribution lines, it is necessary to regularly inspect the transmission and distribution lines, timely detect and eliminate line faults, and improve power supply reliability. Traditional manual inspections have the disadvantages of low efficiency, high risk, and small inspection range, and cannot meet the needs of safe and efficient inspections of transmission and distribution lines.

[0003] In recent years, with the rapid development of drone technology, drones have gradually received high attention from power companies for their high efficiency, low cost, high safety, and easy operation. In the field of power detection, drones, as a new detection platform, have been widely used in line detection. Drones can conduct detailed inspections of transmission and distribution lines, and can accurately identify defects such as hardware shedding and insulator self-explosion that are difficult to find in manual inspections. Drone inspections are not restricted by terrain and are particularly suitable for inspections in dangerous mountainous areas and river areas. At the same time, drone inspections are also suitable for power inspections in some dangerous situations.

[0004] At present, there are problems in the intelligent inspection of power transmission lines by drones, such as single-point flight, disordered inspection, and low fault identification accuracy. Therefore, in order to overcome the shortcomings of the existing technology, the present invention proposes a long-endurance drone intelligent inspection method for transmission and distribution lines. Based on the inspection task, the optimal inspection path is planned, and image data is collected in real time during the inspection process to maintain continuous learning to strengthen the monitoring and control of inspection resources. Summary of the invention

[0005] The purpose of the present invention is to provide a long-endurance drone intelligent inspection method for power transmission and distribution lines. By solving the mathematical model of path planning, the drone inspection plan is obtained. The image data set of the faster RCNN training is used to identify the power transmission and distribution line images collected by the drone. Based on the new image data, continuous learning is maintained to improve the generalization and accuracy of the fault recognition model.

[0006] To achieve the technical purpose, the technical solution of the present invention specifically includes the following steps: Step 1: Considering the location information, quantity information, endurance and other related parameter information of the drones, taking the drone inspection time and total inspection distance as the objective function, constructing the drone intelligent inspection path planning model, and using MOPFA to solve the path planning model to obtain the path planning scheme for each drone; Step 2: Based on the images of power transmission and distribution lines taken by drones during inspection, an image dataset is constructed. To ensure the accuracy of the model, manual intervention can be performed to balance the sample dataset and divide the dataset into a training set and a test set. A fault recognition model is constructed based on the faster CNN algorithm. The model can identify various types of power transmission and distribution line faults. Step 3: Considering that the drone will capture fault types that do not appear in the data set during the continuous inspection process, the fault recognition model has the problem of insufficient adaptability in identifying new faults. The orthogonal weight correction algorithm is used to update the fault recognition model to maintain its continuous learning ability.

[0007] Preferably, the specifics of step 2 are: Step 2-1: Preprocess the historical image dataset, including manual intervention to maintain the balance of samples and divide the training set and test set into fixed proportions; Step 2-2: Input the training set image into the deep convolutional network to obtain the convolution feature map, further use the RPN structure to generate the candidate region, and project it into the feature map to construct the feature matrix; Step 2-3: Use the ROI pooling layer to fix the size of the feature matrix and input it into the fully connected layer; Step 2-4: Calculate the loss function, including classification loss and regression loss, use the joint training method, backpropagate the total loss, and train the Faster RCNN model.

[0008] The beneficial effects of the present invention are: 1. Plan the inspection path of long-endurance drones, take the inspection time and total inspection distance as the optimization goals, consider constraints such as flight conditions and endurance, build a path planning model that conforms to reality, and solve the optimal path based on MOPFA.

[0009] 2. Preprocess the historical image data, establish training sets and test sets, use Faster RCNN to continuously learn the training set, obtain the fault recognition model, and verify the accuracy of the model in the test set. During the inspection process, the drone can input the continuously collected image data into the fault recognition model, and return the recognition results to the control center in real time for the dispatcher to formulate subsequent inspection plans.

[0010] 3. Based on the trained fault recognition model, identify the new images collected during the drone inspection process, maintain continuous learning of the model based on the orthogonal weight correction algorithm, and continuously update the model to improve the accuracy of the fault recognition model. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] The present invention will be further described below in conjunction with the accompanying drawings and implementation examples: Figure 1 This is a flowchart of the intelligent inspection of the long-endurance UAV of the present invention; Figure 2 This is a flow chart of the intelligent inspection path planning of the long-endurance UAV of the present invention; Figure 3 This is a flow chart of fault identification for intelligent inspection of long-endurance UAV of the present invention; DETAILED DESCRIPTION like Figure 1 As shown, the long-endurance drone intelligent inspection method for power transmission and distribution lines disclosed in the present invention includes the following steps: Step 1: Considering the location information, quantity information and relevant parameter information of the UAV’s endurance, taking the inspection time and total inspection distance of the UAV as the objective function, a path planning model is constructed, and the MOPFA algorithm is used to solve the path planning model to obtain the optimal path planning solution for each UAV; Step 2: Based on the images of power transmission and distribution lines taken by drones during inspection, an image dataset is constructed. To ensure the accuracy of the model, manual intervention can be performed to balance the sample dataset and divide the dataset into a training set and a test set. A fault recognition model is constructed based on the faster CNN algorithm. The fault recognition model can identify various types of power transmission and distribution line faults. Step 3: Considering that the UAV will capture fault types that do not appear in the data set during the continuous inspection process, the fault recognition model has the problem of insufficient adaptability in the identification of new faults. The orthogonal weight correction algorithm can maintain the original knowledge while improving the plasticity of subsequent learning, continuously learn new mapping relationships, update the fault recognition model, and maintain the continuous learning ability of the UAV.

[0012] The step 1 is specifically as follows: According to the location information, quantity information and endurance parameter information of the starting point of the UAV, it is converted into the constraint conditions of the path planning model. In order to ensure the high inspection efficiency of the UAV, the inspection time and the total inspection distance are used as the objective function in the path planning model, and the UAV intelligent inspection path planning model is constructed; like Figure 2 To solve the UAV inspection path planning process using MOPFA, the following steps are included: Step 1-1: Randomly initialize the population so that it is evenly distributed in the search area: x=L bd +rand(0,1)*(U bd -L bd )

[0013] Among them, x is the position vector of the individual population, U bd and L bd are the upper and lower boundaries of the search area respectively.

[0014] Step 1-2: Calculate the fitness of each individual, assign the best one as a pathfinder, and store it in memory:

[0015]

[0016] Among them, P i is the probability of deleting the non-dominant solution in the memory by the roulette wheel method, max and min represent the maximum and minimum values ​​of each target, box_size represents the capacity of the memory, and d in is the number of adjacent solutions less than distance d.

[0017] Step 1-3: Leader position update to lead the follower movement:

[0018] Where r1 and u1 are random vectors uniformly generated in the range [0,1], k and k max are the current number of iterations and the maximum number of iterations, respectively, p is the pathfinder position vector.

[0019] Step 1-4: According to the size of the follower, construct the corresponding path-finding matrix. The update process of the follower is as follows:

[0020] Where u2 is a random vector uniformly generated in the range [0,1], x i and x j are the i-th and j-th follower position vectors, respectively, r2 and r3 are uniformly generated random variables in the range [0,1], a is the interaction coefficient, indicating the magnitude of any member moving with its neighbors, and b is the attraction coefficient, which sets the random distance that keeps the followers roughly at a distance from the leader.

[0021] Step 1-5: Determine whether the maximum number of iterations k has been reached max .

[0022] Figure 3 This is a long-endurance UAV intelligent inspection fault identification flow chart, and the step 2 is as follows: First, the image data is input into conv layers, which use VGG to obtain the corresponding feature maps. Secondly, the candidate boxes of each image data are generated based on RPN. Considering that RPN will generate more candidate boxes, the non-maximum suppression algorithm is used to filter the redundancy of the overlapping parts. Specifically, a small network is continuously slid in the convolution feature map to traverse all areas of the feature map and generate candidate boxes.

[0023] Further, it is projected to the corresponding feature map to construct the feature matrix.

[0024] Furthermore, based on the ROI pooling layer, each feature matrix is ​​scaled to the specified size and flattened. The two parallel outputs are the classification layer and the boundary regression network. The classification layer outputs the probability of each category, and the boundary regression network outputs the boundary regression parameters. Back propagation continuously trims the model.

[0025] Further, based on the definition of loss function, including classification loss and boundary regression loss:

[0026]

[0027] In the formula, i is the index of the anchor, p i Represents the predicted probability that the anchor is the target, and the anchor is a positive sample. is 1, and 0 for negative samples, t i Represents the predicted boundary region regression parameter vector of the i-th anchor, A vector of regression parameters representing the true region.

[0028] Furthermore, the joint training method is used to back-propagate the total loss and train the Faster RCNN model.

[0029] Finally, the image data is input and faults are identified in real time.

[0030] In summary, the long-endurance UAV intelligent inspection method for power transmission and distribution lines provided by the present invention plans the path based on the existing UAV intelligent inspection method to improve the inspection efficiency; at the same time, the orthogonal weight algorithm is used to construct an orthogonal projector, which overcomes the catastrophic forgetting of the existing fault recognition model, enhances the recognition accuracy of the model, and has broad application prospects.

Claims

1. A long-endurance drone intelligent inspection method for power transmission and distribution lines, characterized in that: The following steps are involved: Step 1: Taking the inspection time and total inspection distance of the drone as the objective function, a path planning model is constructed, and the MOPFA algorithm is used to solve the path planning model to obtain the optimal path planning solution for the drone; Step 2: Based on the transmission and distribution line images taken by the drone during the inspection, an image dataset is constructed. The image dataset is manually intervened to maintain balance, and the dataset is divided into a training set and a test set. A fault recognition model is constructed based on the faster CNN algorithm. Step 3: The orthogonal weight correction algorithm is used to update the fault recognition model, and the drone maintains continuous learning capabilities. Step 4: The recognition results of the drone inspection process are transmitted to the control center in real time, and the dispatcher will make corresponding response measures and formulate subsequent inspection plans.

2. The long-endurance drone intelligent inspection method for power transmission and distribution lines according to claim 1 is characterized in that: In step 1, the method of using MOPFA to solve the UAV inspection path planning includes the following steps: Step 1-1: Randomly initialize the population so that it is evenly distributed in the search area: x=L bd +rand(0,1)*(U bd -L bd ) Among them, x is the position vector of the individual population, U bd and L bd are the upper and lower boundaries of the search area respectively; Step 1-2: Calculate the fitness of each individual, assign the best one as a pathfinder, and store it in memory: Among them, P i is the probability of deleting the non-dominant solution in the memory by the roulette wheel method, max and min represent the maximum and minimum values ​​of each target, box_size represents the capacity of the memory, and d in is the number of adjacent solutions less than the distance d; Step 1-3: Leader position update to lead the follower movement: Where r1 and u1 are random vectors uniformly generated in the range [0,1], k and k max are the current number of iterations and the maximum number of iterations, respectively, p is the pathfinder position vector; Step 1-4: According to the size of the follower, construct the corresponding path-finding matrix. The update process of the follower is as follows: Where u2 is a random vector uniformly generated in the range [0,1], x i and x j are the position vectors of the ith and jth followers, r2 and r3 are random variables uniformly generated in the range [0,1], a is the interaction coefficient, and b is the attraction coefficient; Step 1-5: Determine whether the maximum number of iterations k has been reached max .

3. The long-endurance drone intelligent inspection method for power transmission and distribution lines according to claim 1 is characterized in that: In step 2, the specific implementation is as follows: Step 2-1: Preprocess the historical image dataset, including manual intervention to maintain the balance of samples and divide the training set and test set into fixed proportions; Step 2-2: Input the training set image into the deep convolutional network to obtain the convolution feature map, further use the RPN structure to generate the candidate region, and project it into the feature map to construct the feature matrix; Step 2-3: Use the ROI pooling layer to fix the size of the feature matrix and input it into the fully connected layer; Step 2-4: Calculate the loss function, including classification loss and regression loss, use the joint training method, backpropagate the total loss, and train the Faster RCNN model.

4. The long-endurance drone intelligent inspection method for power transmission and distribution lines according to claim 1 is characterized in that: In step 2, constraints are constructed on the Faster RCNN fully connected layer based on the orthogonal weight correction algorithm to keep the weight changes in the same feature space, including the initialization stage and the continuous learning stage.

Images