Typical target detection method and system for transmission lines based on UAV federated learning

Abstract

The present invention relates to a typical target detection method and system for power transmission lines based on UAV federated learning. The method comprises: according to a priori typical target database of power transmission lines, using the YOLOv3 target detection algorithm to determine a typical target detection model for power transmission lines; dividing the UAV network into a plurality of federated learning units; obtaining pictures of typical targets of power transmission lines taken by the UAV network and corresponding labels, and determining a training database; according to the training database and the typical target detection model of power transmission lines, training each federated learning unit based on the Horovod architecture and the FATE federated learning architecture, and determining a trained UAV network based on the trained federated learning unit; and using the trained UAV network to determine typical targets in pictures of typical targets of power transmission lines. The present invention can avoid communication link congestion and improve detection efficiency.

Description

Technical Field

[0001] The present invention relates to the field of power transmission line target detection, and in particular to a typical power transmission line target detection method and system based on unmanned aerial vehicle federated learning. Background Art

[0002] Typical targets in transmission lines are an important guarantee for maintaining the normal operation of overhead transmission lines. Take power insulators as an example. They can play a good insulating role in poles, crossarms and power conductors. However, due to long-term exposure to the field, insulators are susceptible to sun, rain and chemical corrosion; on the other hand, due to their special functions and positions in the entire overhead transmission line, insulators are often subjected to strong electric fields and excessive mechanical loads. This also makes insulators very prone to cracking, breakage and lightning flashover. If these faults occur, they will affect the normal operation of the transmission line. Therefore, for the stable operation of overhead transmission lines, it is extremely important to conduct regular inspection and maintenance of typical targets of transmission lines such as power insulators.

[0003] Most traditional inspection methods rely on manual inspections, but manual inspections have the disadvantages of being time-consuming, high in labor costs, high in work risks, and low in efficiency. In recent years, with the rapid progress of computer vision technology and drone technology, the use of drone aerial images and intelligent recognition and positioning algorithms to replace manual inspections has become one of the mainstream development trends. By observing the transmission line images taken by drones, the status of various transmission line equipment can be visually checked, such as whether the wires are covered with ice and broken, whether the pins are loose, and whether the insulators are cracked, and then staff can be dispatched to troubleshoot. However, the traditional distributed model training method will cause communication link congestion due to the centralized upload of data, resulting in low detection efficiency. Summary of the invention

[0004] The purpose of the present invention is to provide a typical target detection method and system for power transmission lines based on UAV federated learning, which can avoid communication link congestion and improve detection efficiency.

[0005] To achieve the above object, the present invention provides the following solutions:

[0006] A typical target detection method for power transmission lines based on UAV federated learning, including:

[0007] According to a priori database of typical targets of transmission lines, a typical target detection model of transmission lines is determined by using the YOLOv3 target detection algorithm; the priori database of typical targets of transmission lines includes: pictures of typical targets of transmission lines and corresponding labels; the typical targets of transmission lines include: insulators, conductors or pins;

[0008] Divide the UAV network into multiple federated learning units; obtain pictures of typical targets of power transmission lines taken by the UAV network and their corresponding labels, and determine the training database;

[0009] According to the training database and the typical target detection model of the power transmission line, each federated learning unit is trained based on the Horovod architecture and the FATE federated learning architecture, and the trained UAV network is determined based on the trained federated learning units;

[0010] Use the trained drone network to identify typical targets in images of typical targets of power transmission lines.

[0011] Optionally, the training of each federated learning unit based on the Horovod architecture and the FATE federated learning architecture according to the training database and the typical target detection model of the transmission line, and determining the trained UAV network according to the trained federated learning unit specifically includes:

[0012] According to the training database and the typical target detection model of the transmission line, the parallel computing power of each federated learning unit is allocated based on the Horovod architecture, and the weight of each federated learning unit is determined;

[0013] Use the FATE federated learning architecture to aggregate the weights of each federated learning unit;

[0014] The aggregated weights are passed to each federated learning unit, and the steps of allocating the parallel computing power of each federated learning unit based on the Horovod architecture according to the training database and the typical target detection model of the transmission line to determine the weight of each federated learning unit are returned until the Loss error function converges to determine the trained UAV network.

[0015] Optionally, the Loss error function is:

[0016] Loss = λ coord L1+λ coord L2+L3+L4;

[0017] Among them, L1 is the center coordinate error, L2 is the width and height coordinate error, L3 is the confidence error, L4 is the classification error, and λcoord is the joint error coefficient.

[0018] A typical target detection method for power transmission lines based on UAV federated learning, including:

[0019] A transmission line typical target detection model determination module is used to determine a transmission line typical target detection model using a YOLOv3 target detection algorithm based on a transmission line typical target prior database; the transmission line typical target prior database includes: images of transmission line typical targets and corresponding labels; the transmission line typical targets include: insulators, wires or pins;

[0020] The module for determining the federated learning unit and the training database is used to divide the UAV network into multiple federated learning units; obtain the pictures of typical targets of the power transmission line taken by the UAV network and the corresponding labels, and determine the training database;

[0021] A trained UAV network determination module is used to train each federated learning unit based on the Horovod architecture and the FATE federated learning architecture according to the training database and the typical target detection model of the transmission line, and determine the trained UAV network according to the trained federated learning unit;

[0022] The detection module is used to use the trained UAV network to determine typical targets in the picture of typical targets of power transmission lines.

[0023] Optionally, the trained drone network determination module specifically includes:

[0024] The weight training unit of the federated learning unit is used to allocate the parallel computing power of each federated learning unit based on the Horovod architecture according to the training database and the typical target detection model of the transmission line, and determine the weight of each federated learning unit;

[0025] The weight aggregation unit of the federated learning unit is used to aggregate the weight of each federated learning unit using the FATE federated learning architecture;

[0026] The iteration unit is used to transfer the summarized weights to each federated learning unit and return the weight training unit of the federated learning unit.

[0027] Optionally, the Loss error function is:

[0028] Loss = λ coord L1+λ coord L2+L3+L4;

[0029] Among them, L1 is the center coordinate error, L2 is the width and height coordinate error, L3 is the confidence error, L4 is the classification error, and λcoord is the joint error coefficient.

[0030] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0031] The present invention provides a typical target detection method and system for power transmission lines based on UAV federated learning. The method of taking target pictures by UAV inspection replaces the traditional manual power transmission line inspection, which can effectively solve the shortcomings of low efficiency and high resource consumption of manual inspection. Compared with the traditional distributed model training method completed on a server unit with simple connection, the UAV federated learning architecture can efficiently complete the update and feedback of training parameters, without the need to centrally upload data to avoid communication link congestion, and can maximize the privacy of data of each federated learning unit. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0033] Figure 1 A schematic diagram of a typical target detection method for power transmission lines based on UAV federated learning provided by the present invention;

[0034] Figure 2 Flowchart of the federated learning process for drones;

[0035] Figure 3 Schematic diagram of the UAV federated learning process;

[0036] Figure 4 This is a graph of the Loss error function during the UAV federated learning process;

[0037] Figure 5 This is a schematic diagram of the structure of a typical target detection system for power transmission lines based on UAV federated learning provided by the present invention. DETAILED DESCRIPTION

[0038] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0039] The purpose of the present invention is to provide a typical target detection method and system for power transmission lines based on UAV federated learning, which can avoid communication link congestion and improve detection efficiency.

[0040] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0041] Figure 1 This is a flow chart of a typical target detection method for power transmission lines based on UAV federated learning provided by the present invention. Figure 2 This paper proposes a typical target detection method for power transmission lines based on UAV federated learning, including:

[0042] S101, according to a priori database of typical targets of transmission lines I0, a typical target detection model of transmission lines is determined by using a YOLOv3 target detection algorithm, that is, a weight P0 of the model is determined; the priori database of typical targets of transmission lines includes: pictures of typical targets of transmission lines and corresponding labels; the typical targets of transmission lines include: insulators, conductors or pins;

[0043] S102, dividing the UAV network into multiple federated learning units; obtaining pictures of typical targets of the power transmission line taken by the UAV network and corresponding labels, and determining a training database;

[0044] Press each i The UAV network containing N UAV nodes is divided into l federated learning units in the way that N UAVs form a federated learning unit.

[0045] Parameters ai, N, l satisfy:

[0046]

[0047] Where a i Represents the number of drone nodes in the i-th federated learning unit.

[0048] The number of photos M taken by the drone satisfies:

[0049]

[0050] Where L represents the total number of training labels marked by M pictures, b j Represents the number of training tags in the jth image, Tag jk represents the kth training label in the jth image.

[0051] S103, according to the training database and the typical target detection model of the transmission line, each federated learning unit is trained based on the Horovod architecture and the FATE federated learning architecture, and the trained UAV network is determined according to the trained federated learning units;

[0052] Specifically include:

[0053] According to the training database and the typical target detection model of the transmission line, the parallel computing power of each federated learning unit is allocated based on the Horovod architecture, and the weight of each federated learning unit is determined;

[0054] Use the FATE federated learning architecture to aggregate the weights of each federated learning unit;

[0055] The aggregated weights are passed to each federated learning unit, and the steps of allocating the parallel computing power of each federated learning unit based on the Horovod architecture according to the training database and the typical target detection model of the transmission line to determine the weight of each federated learning unit are returned until the Loss error function converges to determine the trained UAV network.

[0056] The Loss error function is:

[0057] Loss = λ coord L1+λ coord L2+L3+L4;

[0058] Among them, L1 is the center coordinate error, L2 is the width and height coordinate error, L3 is the confidence error, L4 is the classification error, and λcoord is the joint error coefficient.

[0059] like Figure 2 and Figure 3 As shown in the figure, Horovod is a distributed training framework that uses the Ring-Allreduce algorithm. The basic process of distributed training using the Ring-Allreduce algorithm is as follows:

[0060] Step S1: Each UAV device calculates the gradient according to its own training data to obtain the gradient;

[0061] Step S2: Divide the gradient vector on each UAV device into H slices of roughly equal length (where the number of slices H is equal to the number of devices);

[0062] Step S3: After H-1 rounds of gradient transmission and gradient addition, a small part of the gradient vector on each UAV device is the sum of the gradients of the slice in all devices;

[0063] Step S4: broadcast the sum of each gradient vector slice calculated in the previous stage to other UAV devices through H-1 rounds of gradient transmission;

[0064] Step S5: Merge the shard gradients on each drone device and update the model on each device based on the gradients.

[0065] Assume that the total number of parameters of the trained model is X, the number of Parameter Server units in the model is 1, and the number of Worker units is H.

[0066] Based on this, the number of times each drone device in the Horovod framework needs to transmit information is:

[0067] E = 2N-1;

[0068] The time T required for each batchinteration data transmission of the training model is:

[0069]

[0070] As N gradually increases, T approaches T′, which satisfies:

[0071]

[0072] The functions of each module in the FATE federated learning architecture are:

[0073] FATE Flow: This service is divided into two parts: Client and Server. The Client part is used by users to submit federated learning tasks to the FATE cluster; FATE Flow Server is the entrance for the FATE cluster to provide services to the outside world;

[0074] MySQL: Some metadata related to the task, such as creation time and status, will be stored in MySQL.

[0075] EGG / ROLL: Provides distributed computing and storage capabilities for training tasks.

[0076] Meta Service: A set of data or a file can be sliced ​​and distributed on different eggs. Metaservice is responsible for managing and locating the slice information of the file.

[0077] Federation: Due to the special nature of federated learning, data is often exchanged several times between participants during training. This service provides the function of sending and receiving data for training tasks.

[0078] Proxy: Provides reverse proxy service and is the only entrance for the FATE cluster to the outside world (training other participants).

[0079] FATE Board: Provides visualization of training tasks to users.

[0080] FATE Serving: Provides online reasoning services. Users can push trained models to the service for online reasoning.

[0081] The following example illustrates the process of federated learning by two different UAV federated learning units A and B under the coordination of collaborator C:

[0082] Step S1: Based on the encrypted user sample alignment technology, A and B confirm the shared image data of both parties without disclosing their respective data, so as to combine the features of these image data for modeling;

[0083] Step S2: Collaborator C distributes the public key to A and B to encrypt the data that needs to be exchanged during the training process;

[0084] Step S3: A and B respectively calculate the intermediate results of the gradient and send the results to C in encrypted form. At the same time, B calculates the loss function value based on its label data;

[0085] Step S4: C decrypts the results sent by A and B, calculates the total gradient value, and then encrypts the total gradient value and sends it to A and B.

[0086] Step S5: A and B update the parameters of their respective models according to the gradient and start a new round of training process.

[0087] Figure 4 It reflects the convergence process of the Loss error function value in 1-100 rounds of model iterative training. The analytical formula of the Loss error function is:

[0088] Loss = λ coord L1+λ coord L2+L3+L4

[0089] Where L1 is the center coordinate error, L2 is the width and height coordinate error, L3 is the confidence error, and L4 is the classification error.

[0090] The center coordinate error L1 satisfies:

[0091]

[0092] Where S 2 Indicates that the number of grids is S×S, B indicates the number of each candidate box, Indicates the responsibility of the j-th candidate box of the p-th grid for the project. and They represent the predicted value and true value of the horizontal coordinate of the center point of the j-th candidate box of the p-th grid, respectively. and They respectively represent the predicted value and true value of the ordinate of the center point of the j-th candidate box in the p-th grid.

[0093] The width and height coordinate error L2 satisfies:

[0094]

[0095] In the formula and They represent the predicted value and true value of the width of the j-th candidate box of the p-th grid, respectively. and They represent the predicted value and true value of the height of the j-th candidate box in the p-th grid respectively.

[0096] The confidence error L3 satisfies:

[0097]

[0098] In the formula Indicates the irresponsibility of the j-th candidate box of the p-th grid to the project. and Represents the predicted value and true value of the confidence of the j-th candidate box in the p-th grid.

[0099] The classification error L4 satisfies:

[0100]

[0101] In the formula and They represent the predicted value and true value of the classification probability of the j-th candidate box in the p-th grid respectively.

[0102] S104, using the trained UAV network to determine typical targets in the picture of typical targets of the power transmission line.

[0103] Figure 5 This is a schematic diagram of a typical target detection system for power transmission lines based on UAV federated learning provided by the present invention. Figure 5 As shown, the present invention provides a typical target detection method for power transmission lines based on UAV federated learning, including:

[0104] The transmission line typical target detection model determination module 501 is used to determine the transmission line typical target detection model using the YOLOv3 target detection algorithm according to the transmission line typical target prior database; the transmission line typical target prior database includes: pictures of transmission line typical targets and corresponding labels; the transmission line typical targets include: insulators, wires or pins;

[0105] The federated learning unit and training database determination module 502 is used to divide the UAV network into multiple federated learning units; obtain pictures of typical targets of the power transmission line taken by the UAV network and corresponding labels, and determine the training database;

[0106] A trained UAV network determination module 503 is used to train each federated learning unit based on the Horovod architecture and the FATE federated learning architecture according to the training database and the typical target detection model of the transmission line, and determine the trained UAV network according to the trained federated learning unit;

[0107] The detection module 504 is used to use the trained UAV network to determine typical targets in the picture of typical targets of the power transmission line.

[0108] The trained UAV network determination module 503 specifically includes:

[0109] The weight training unit of the federated learning unit is used to allocate the parallel computing power of each federated learning unit based on the Horovod architecture according to the training database and the typical target detection model of the transmission line, and determine the weight of each federated learning unit;

[0110] The weight aggregation unit of the federated learning unit is used to aggregate the weight of each federated learning unit using the FATE federated learning architecture;

[0111] The iteration unit is used to transfer the summarized weights to each federated learning unit and return the weight training unit of the federated learning unit.

[0112] The Loss error function is:

[0113] Loss = λ coord L1+λ coord L2+L3+L4;

[0114] Among them, L1 is the center coordinate error, L2 is the width and height coordinate error, L3 is the confidence error, L4 is the classification error, and λ coord is the joint error coefficient.

[0115] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the system disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0116] The principles and implementation methods of the present invention are described in this article using specific examples. The description of the above embodiments is only used to help understand the method and core idea of ​​the present invention. At the same time, for those skilled in the art, according to the idea of ​​the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting the present invention.

Claims

1. A typical target detection method for power transmission lines based on UAV federated learning, characterized in that: include: According to the prior database of typical targets of transmission lines, the YOLOv3 target detection algorithm is used to determine the typical target detection model of transmission lines; The priori database of typical targets of transmission lines includes: pictures of typical targets of transmission lines and corresponding labels; the typical targets of transmission lines include: insulators, conductors or pins; Divide the UAV network into multiple federated learning units; obtain pictures of typical targets of power transmission lines taken by the UAV network and their corresponding labels, and determine the training database; According to the training database and the typical target detection model of the power transmission line, each federated learning unit is trained based on the Horovod architecture and the FATE federated learning architecture, and the trained UAV network is determined based on the trained federated learning units; Use the trained UAV network to identify typical targets in the image of typical targets of power transmission lines; The method trains each federated learning unit based on the Horovod architecture and the FATE federated learning architecture according to the training database and the typical target detection model of the transmission line, and determines the trained drone network according to the trained federated learning unit, specifically including: According to the training database and the typical target detection model of the transmission line, the parallel computing power of each federated learning unit is allocated based on the Horovod architecture, and the weight of each federated learning unit is determined; Use the FATE federated learning architecture to aggregate the weights of each federated learning unit; The aggregated weights are passed to each federated learning unit, and the steps of allocating the parallel computing power of each federated learning unit based on the Horovod architecture according to the training database and the typical target detection model of the transmission line to determine the weight of each federated learning unit are returned until the Loss error function converges to determine the trained UAV network.

2. According to claim 1, a typical target detection method for transmission lines based on UAV federated learning is characterized in that: The Loss error function is: Loss=λ coord L1+λ coord L2+L3+L4; Among them, L1 is the center coordinate error, L2 is the width and height coordinate error, L3 is the confidence error, L4 is the classification error, and λcoord is the joint error coefficient.

3. A transmission line typical target detection system based on UAV federated learning, used to implement a transmission line typical target detection method based on UAV federated learning as described in any one of claims 1-2, characterized in that: include: A transmission line typical target detection model determination module is used to determine a transmission line typical target detection model using a YOLOv3 target detection algorithm based on a transmission line typical target priori database; The priori database of typical targets of transmission lines includes: pictures of typical targets of transmission lines and corresponding labels; the typical targets of transmission lines include: insulators, conductors or pins; The module for determining the federated learning unit and the training database is used to divide the UAV network into multiple federated learning units; obtain the pictures of typical targets of the power transmission line taken by the UAV network and the corresponding labels, and determine the training database; A trained UAV network determination module is used to train each federated learning unit based on the Horovod architecture and the FATE federated learning architecture according to the training database and the typical target detection model of the transmission line, and determine the trained UAV network according to the trained federated learning unit; The detection module is used to use the trained UAV network to determine typical targets in the picture of typical targets of power transmission lines.

4. A typical target detection system for power transmission lines based on UAV federated learning according to claim 3, characterized in that: The trained UAV network determination module specifically includes: The weight training unit of the federated learning unit is used to allocate the parallel computing power of each federated learning unit based on the Horovod architecture according to the training database and the typical target detection model of the transmission line, and determine the weight of each federated learning unit; The weight aggregation unit of the federated learning unit is used to aggregate the weight of each federated learning unit using the FATE federated learning architecture; The iteration unit is used to transfer the summarized weights to each federated learning unit and return the weight training unit of the federated learning unit.

5. A typical target detection system for power transmission lines based on UAV federated learning according to claim 4, characterized in that Lo The ss error function is: Loss=λ coord L1+λ coord L2+L3+L4; Among them, L1 is the center coordinate error, L2 is the width and height coordinate error, L3 is the confidence error, L4 is the classification error, and λcoord is the joint error coefficient.

Images