Model compression analysis method for intelligent identification of foreign objects in power transmission corridors

Abstract

The present invention provides a model compression analysis method for intelligent identification of foreign objects in a transmission corridor, comprising the steps of: S101, training a first neural network model based on a training set of foreign object images in a transmission corridor on a cloud server, compressing the first neural network model after the training is completed to obtain a second neural network model; S102, deploying the second neural network model to an edge computing terminal; S103, collecting the transmission corridor image through an image acquisition device, and transmitting the transmission corridor image to the edge computing terminal; S104, the edge computing terminal inputs the transmission corridor image into the second neural network model, performs foreign object identification on the transmission corridor image through the second neural network model, sends the foreign object identification result to an operation and maintenance terminal, and uploads it to the cloud server. The method can analyze and identify foreign objects in transmission corridors in remote areas in real time and notify operation and maintenance personnel in a timely manner, thereby improving the efficiency of handling foreign objects in transmission corridors and better ensuring the safe and stable operation of transmission lines.

Description

Technical Field

[0001] The present invention relates to the technical field of foreign body identification in power transmission corridors, and in particular to a model compression analysis method for intelligent identification of foreign bodies in power transmission corridors. Background Art

[0002] Since electricity plays an indispensable role in the rapid development of social economy, higher requirements must be put forward for the safety of power transmission. As an important part of power transmission, the maintenance of transmission lines is the basis for ensuring the safety of power transmission. Transmission lines are prone to damage caused by foreign objects such as kites, agricultural greenhouses, balloons and advertising banners. In extremely severe weather, if foreign objects are hanging on transmission lines, it is easy to cause power accidents such as fire and short circuit, leading to power outages, and then causing economic losses. In order to improve the efficiency of inspection, in the field of transmission line inspection, intelligent equipment is currently used to collect on-site data information such as pictures or videos and transmit them to the background, and then further identify foreign objects in the transmission line in the image through neural networks. Although neural networks perform well in computer vision tasks, this is mainly due to the increasing network depth and massive trainable parameters. Although the recognition accuracy is improved, the hardware performance requirements of the server are also increased accordingly. When transmission lines are deployed in remote areas, on-site image data needs to be remotely transmitted back to the server for recognition, and real-time analysis cannot be achieved. For this, edge computing devices can be used to recognize and process images locally, but the large-scale neural network places high requirements on the performance of edge computing devices. Summary of the invention

[0003] In view of this, the purpose of the present invention is to provide a model compression analysis method for intelligent identification of foreign objects in transmission corridors. By compressing the size and parameters of the neural network model, it can be deployed on an edge computing device to achieve real-time analysis and identification of transmission corridor images.

[0004] To achieve the above-mentioned object of the invention, the present invention provides a model compression analysis method for intelligent identification of foreign objects in a transmission corridor, the method comprising the following steps:

[0005] S101, training a first neural network model based on a training set of foreign body images in a power transmission corridor on a cloud server, and compressing the first neural network model after the training is completed to obtain a second neural network model;

[0006] S102, deploying the second neural network model to the edge computing terminal;

[0007] S103, collecting the transmission corridor image through an image acquisition device, and transmitting the transmission corridor image to an edge computing terminal;

[0008] S104. The edge computing terminal inputs the transmission corridor image into the second neural network model, identifies foreign objects in the transmission corridor image through the second neural network model, sends the foreign object identification result to the operation and maintenance terminal, and uploads it to the cloud server.

[0009] Furthermore, the first neural network model is compressed, specifically comprising the following steps:

[0010] S201, creating a compression strategy library, and writing multiple model compression strategies into the compression strategy library;

[0011] S202: Select a model compression strategy from a compression strategy library, and perform compression processing on the first neural network model based on the selected model compression strategy.

[0012] Furthermore, in step S202, a model compression strategy is selected from a compression strategy library, which specifically includes the following steps:

[0013] S301, respectively use each model compression strategy in the compression strategy library to simulate compression of the first neural network model, and obtain a corresponding compressed neural network model as a simulation compression result;

[0014] S302, performing a performance test on the compressed neural network model in the simulation compression result, and evaluating the performance of the compressed neural network according to the performance test result;

[0015] S303: Select and output the model compression strategy corresponding to the simulation compression result with the best performance evaluation.

[0016] Furthermore, before executing step S102, the following steps are also executed:

[0017] S401, the edge computing terminal aggregates the transmission corridor images uploaded by the image acquisition device plugged in thereto, and uploads the transmission corridor images to the cloud server;

[0018] S402, the cloud server inputs the transmission corridor image uploaded by each edge computing terminal into the clustering algorithm, and performs cluster analysis on the edge computing terminals based on the transmission corridor image through the clustering algorithm to obtain multiple clusters, each cluster containing multiple edge computing terminals;

[0019] S403: Determine a representative transmission corridor image for each cluster.

[0020] Further, step S202 specifically includes the following steps:

[0021] S501, summarizing representative transmission corridor images of all clusters to generate an image set;

[0022] S502, using each model compression strategy in the compression strategy library to simulate compression of the first neural network model, and obtaining a corresponding compressed neural network model as a simulation compression result;

[0023] S503, generating a compression result set, where the compression result set includes all simulation compression results;

[0024] S504, inputting the image set and the simulation compression result set into a genetic algorithm, setting the goal of the genetic algorithm to have a foreign body recognition accuracy higher than a preset accuracy threshold while minimizing the model size, performing calculations through the genetic algorithm, and outputting the optimal individual;

[0025] S505. Deploy the corresponding second neural network model to the edge computing terminal based on the optimal individual.

[0026] Furthermore, step S504 specifically includes the following steps:

[0027] S601, generating a plurality of individuals based on an image set and a simulation compression result set, and using the plurality of individuals as an initial population of a genetic algorithm;

[0028] S602, randomly selecting individuals from the initial population and outputting the selected individuals;

[0029] S603, performing mutation and crossover operations on the selected individuals;

[0030] S604, iteratively execute steps S602 to S603 until the genetic algorithm converges and outputs the population.

[0031] Further, multiple individuals are generated based on the image set and the simulation compression result set, specifically including the following steps:

[0032] S701, traversing the image set, randomly selecting a simulation compression result from the simulation compression result set for each representative power transmission corridor image in the image set for binding, and forming the selected simulation compression results into a second simulation compression result set;

[0033] S702, taking the paired image set and the second simulation compression result set as an individual of the population;

[0034] S703. Repeat steps S701 to S702 until the number of individuals generated meets the population quantity requirement.

[0035] Further, the second neural network model is deployed to the edge computing terminal, specifically including the steps of:

[0036] S801, parsing the image set and simulation compression result set contained in the optimal individual, and determining different representative transmission corridor images in the optimal individual and their corresponding simulation compression results;

[0037] S802, determining a corresponding model compression strategy according to the simulation compression result, and compressing the first neural network model based on the corresponding model compression strategy to obtain a corresponding second neural network model;

[0038] S803. Deploy the second neural network model obtained in the previous step to the edge computing terminal in the cluster corresponding to the representative transmission corridor image bound to the simulation compression result.

[0039] Compared with the prior art, the present invention has the following beneficial effects:

[0040] The present invention provides a model compression and analysis method for intelligent identification of foreign objects in a transmission corridor. After a cloud server trains a first neural network model, it is further compressed to obtain a second neural network model with a smaller volume and fewer parameters, and the second neural network model is deployed to an edge computing terminal. An image acquisition device acquires an image of the transmission corridor and transmits it to the edge computing terminal. The edge computing terminal identifies the image of the transmission corridor and sends the foreign object identification result to the operation and maintenance terminal, thereby reducing the hardware performance requirements of the neural network model for the edge computing terminal and reducing the occupancy of cloud server computing resources. The method can analyze and identify foreign objects in transmission corridors in remote areas in real time and notify operation and maintenance personnel in a timely manner, thereby improving the efficiency of handling foreign objects in transmission corridors and better ensuring the safe and stable operation of transmission lines. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only preferred embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0042] Figure 1 It is a schematic diagram of the overall process of a model compression analysis method for intelligent identification of foreign objects in a transmission corridor provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0043] The principles and features of the present invention are described below in conjunction with the accompanying drawings. The enumerated embodiments are only used to explain the present invention and are not used to limit the scope of the present invention.

[0044] Reference Figure 1 This embodiment provides a model compression analysis method for intelligent identification of foreign objects in power transmission corridors, the method comprising the following steps:

[0045] S101. Train the first neural network model based on the transmission corridor foreign body image training set on the cloud server, and compress the first neural network model after the training is completed to obtain a second neural network model. The first neural network model can be Faster R-CNN, or a YOLOv3 network, or other networks that can realize foreign body image recognition, which is not specifically limited in this embodiment. The second neural network model is a neural network model with a smaller network scale and smaller parameters after compressing the first neural network model.

[0046] S102: deploy the second neural network model to the edge computing terminal.

[0047] S103, collecting images of the transmission corridor through an image acquisition device, and transmitting the images of the transmission corridor to an edge computing terminal. In this embodiment, multiple image acquisition devices are deployed in the transmission corridor, and each edge computing terminal is externally connected with multiple image acquisition devices in a certain area. The image acquisition devices collect images of the transmission corridor in real time and transmit them to the edge computing terminal.

[0048] S104, the edge computing terminal inputs the transmission corridor image into the second neural network model, performs foreign object recognition on the transmission corridor image through the second neural network model, sends the foreign object recognition result to the operation and maintenance terminal, and uploads it to the cloud server. The operation and maintenance terminal is held by the operation and maintenance personnel, which can be a smart phone, a handheld computer or a computer device. By sending the foreign object recognition result to the operation and maintenance terminal, the operation and maintenance personnel can understand whether there are foreign objects hanging in the transmission corridor and deal with them in time.

[0049] In step S101, the first neural network model is compressed, which specifically includes the following steps:

[0050] S201, create a compression strategy library, and write multiple model compression strategies into the compression strategy library. Each model compression strategy corresponds to a method for compressing a neural network model, such as model pruning, model quantization, model distillation, NAS space search, etc.

[0051] S202, selecting a model compression strategy from a compression strategy library, and performing compression processing on the first neural network model based on the selected model compression strategy.

[0052] As a preferred example, in step S202, selecting a model compression strategy from a compression strategy library specifically includes the following steps:

[0053] S301. Use each model compression strategy in the compression strategy library to simulate compression of the first neural network model, and obtain the corresponding compressed neural network model as a simulation compression result.

[0054] S302, performing a performance test on the compressed neural network model in the simulation compression result, and evaluating the performance of the compressed neural network based on the performance test result.

[0055] S303: Select and output the model compression strategy corresponding to the simulation compression result with the best performance evaluation.

[0056] Since the transmission corridor spans a wide area, the images collected by image acquisition devices deployed at different locations will have different characteristics. It is difficult to further improve the recognition accuracy by uniformly processing these images with different characteristics through a single-structure neural network model. In order to solve this problem, in another preferred example, before executing step S102, the following steps are also performed:

[0057] S401. The edge computing terminal aggregates the transmission corridor images uploaded by the image acquisition device plugged in thereto, and uploads the transmission corridor images to the cloud server.

[0058] S402, the cloud server inputs the transmission corridor images uploaded by each edge computing terminal into the clustering algorithm, and performs cluster analysis on the edge computing terminals based on the transmission corridor images through the clustering algorithm to obtain multiple clusters, each of which contains multiple edge computing terminals. This step performs cluster analysis based on the transmission corridor image features uploaded by each edge computing terminal, and divides the edge computing terminals with the same or similar collected transmission corridor image features into the same cluster, so as to facilitate the subsequent deployment of different neural network models according to the cluster to which the edge computing terminal belongs.

[0059] S403: Determine a representative transmission corridor image for each cluster. The representative transmission corridor image is selected from the transmission corridor images uploaded by the edge computing terminals in the cluster, and the representative transmission corridor image should best reflect the relevant image features of the transmission corridor images uploaded by the edge computing terminals in the cluster to which it belongs.

[0060] Based on the above embodiment, step S202 specifically includes the following steps:

[0061] S501. Summarize representative transmission corridor images of all clusters to generate an image set.

[0062] S502. Use each model compression strategy in the compression strategy library to simulate compression of the first neural network model, and obtain a corresponding compressed neural network model as a simulation compression result.

[0063] S503: Generate a compression result set, where the compression result set includes all simulation compression results.

[0064] S504, input the image set and the simulation compression result set into the genetic algorithm, set the goal of the genetic algorithm to have a foreign body recognition accuracy higher than a preset accuracy threshold while minimizing the model size, calculate through the genetic algorithm, and output the optimal individual.

[0065] S505. Deploy the corresponding second neural network model to the edge computing terminal based on the optimal individual.

[0066] Wherein, step S504 specifically includes the following steps:

[0067] S601, generating multiple individuals based on an image set and a simulation compression result set, and using the multiple individuals as an initial population of a genetic algorithm.

[0068] As a preferred example, generating multiple individuals based on the image set and the simulation compression result set specifically includes the following steps:

[0069] S701, traversing the image set, randomly selecting a simulation compression result from the simulation compression result set for each representative transmission corridor image in the image set to bind, and forming the selected simulation compression results into a second simulation compression result set.

[0070] S702: Take the paired image set and the second simulation compression result set as an individual of the population.

[0071] S703. Repeat steps S701 to S702 until the number of individuals generated meets the population quantity requirement.

[0072] S602: Randomly select individuals from the initial population and output the selected individuals.

[0073] S603. Perform mutation and crossover operations on the selected individuals.

[0074] S604, iteratively execute steps S602 to S603 until the genetic algorithm converges and outputs the population.

[0075] Based on the above embodiment, the second neural network model is deployed to the edge computing terminal, specifically including the steps of:

[0076] S801. Analyze the image set and simulation compression result set included in the optimal individual, and determine different representative transmission corridor images in the optimal individual and their corresponding simulation compression results.

[0077] S802. Determine a corresponding model compression strategy according to the simulation compression result, and compress the first neural network model based on the corresponding model compression strategy to obtain a corresponding second neural network model.

[0078] S803. Deploy the second neural network model obtained in the previous step to the edge computing terminal in the cluster corresponding to the representative transmission corridor image bound to the simulation compression result.

[0079] The method provided in this embodiment first performs cluster analysis on the edge computing terminals according to the transmission corridor images uploaded by each edge computing terminal, and divides the edge computing terminals with the same or similar collected image features into the same cluster, and then inputs the representative transmission corridor images of each cluster and the simulated compression results of different model compression strategies into the genetic algorithm, and selects the optimal individual with the goal of foreign object recognition accuracy higher than a preset accuracy threshold and minimum model size, and determines the second neural network model applicable to the edge computing terminal in the cluster based on the simulated compression results corresponding to the representative transmission corridor images in the optimal individual, and deploys the corresponding second neural network model to the corresponding edge computing terminal, so that the second neural network models deployed on different edge computing terminals are compressed and processed specifically for the collected transmission corridor image features, which can ensure that the foreign object recognition accuracy is higher than the preset accuracy threshold, and the model size is minimized, thereby reducing the requirements on the computing power of the edge computing terminal.

[0080] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A model compression analysis method for intelligent identification of foreign objects in power transmission corridors, characterized in that: The method comprises the following steps: S101, training a first neural network model based on a training set of foreign body images in a power transmission corridor on a cloud server, and compressing the first neural network model after the training is completed to obtain a second neural network model; S102, deploying the second neural network model to the edge computing terminal; S103, collecting the transmission corridor image through an image acquisition device, and transmitting the transmission corridor image to an edge computing terminal; S104, the edge computing terminal inputs the transmission corridor image into the second neural network model, performs foreign object recognition on the transmission corridor image through the second neural network model, sends the foreign object recognition result to the operation and maintenance terminal, and uploads it to the cloud server; The first neural network model is compressed, specifically comprising the following steps: S201, creating a compression strategy library, and writing a plurality of model compression strategies into the compression strategy library; S202, selecting a model compression strategy from a compression strategy library, and compressing the first neural network model based on the selected model compression strategy; In step S202, a model compression strategy is selected from the compression strategy library, which specifically includes the following steps: S301, using each model compression strategy in the compression strategy library to simulate compression of the first neural network model, and obtaining a corresponding compressed neural network model as a simulation compression result; S302, performing a performance test on the compressed neural network model in the simulation compression result, and evaluating the performance of the compressed neural network according to the performance test result; S303, selecting and outputting the model compression strategy corresponding to the simulation compression result with the best performance evaluation; Before executing step S102, the following steps are further executed: S401, the edge computing terminal aggregates the transmission corridor images uploaded by the image acquisition device plugged therein, and uploads the transmission corridor images to the cloud server; S402, the cloud server inputs the transmission corridor image uploaded by each edge computing terminal into the clustering algorithm, and performs cluster analysis on the edge computing terminals based on the transmission corridor image through the clustering algorithm to obtain multiple clusters, each cluster containing multiple edge computing terminals; S403, determining a representative transmission corridor image for each cluster; Step S202 specifically includes the following steps: S501, summarizing representative transmission corridor images of all clusters to generate an image set; S502, using each model compression strategy in the compression strategy library to simulate compression of the first neural network model, and obtaining a corresponding compressed neural network model as a simulation compression result; S503, generating a compression result set, where the compression result set includes all simulation compression results; S504, inputting the image set and the simulation compression result set into a genetic algorithm, setting the goal of the genetic algorithm to have a foreign body recognition accuracy higher than a preset accuracy threshold while minimizing the model size, performing calculations through the genetic algorithm, and outputting the optimal individual; S505. Deploy the corresponding second neural network model to the edge computing terminal based on the optimal individual.

2. According to claim 1, a model compression analysis method for intelligent identification of foreign objects in power transmission corridors is characterized in that: Step S504 specifically includes the following steps: S601, generating multiple individuals based on the image set and the simulation compression result set, and using the multiple individuals as the initial population of the genetic algorithm; S602, randomly selecting individuals from the initial population and outputting the selected individuals; S603, performing mutation and crossover operations on the selected individuals; S604, iteratively execute steps S602 to S603 until the genetic algorithm converges and outputs the population.

3. According to claim 2, a model compression analysis method for intelligent identification of foreign objects in power transmission corridors is characterized in that: Generate multiple individuals based on the image set and the simulation compression result set, specifically including the following steps: S701, traverse the image set, randomly select a simulation compression result from the simulation compression result set for each representative transmission corridor image in the image set to bind, and form the selected simulation compression results into a second simulation compression result set; S702, taking the paired image set and the second simulation compression result set as an individual of the population; S703. Repeat steps S701 to S702 until the number of individuals generated meets the population quantity requirement.

4. According to claim 3, a model compression analysis method for intelligent identification of foreign objects in power transmission corridors is characterized in that: The second neural network model is deployed to the edge computing terminal, specifically including the steps of: S801, parsing the image set and the simulation compression result set contained in the optimal individual, determining different representative transmission corridor images in the optimal individual and their corresponding simulation compression results; S802, determining a corresponding model compression strategy according to the simulation compression result, and compressing the first neural network model based on the corresponding model compression strategy to obtain a corresponding second neural network model; S803. Deploy the second neural network model obtained in the previous step to the edge computing terminal in the cluster corresponding to the representative transmission corridor image bound to the simulation compression result.

Images